blob: b934223eaa456aa664cfd5eb07cab634b70e0386 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/vmscan.c
3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 *
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
12 */
13
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -070014#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
15
Linus Torvalds1da177e2005-04-16 15:20:36 -070016#include <linux/mm.h>
17#include <linux/module.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090018#include <linux/gfp.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070019#include <linux/kernel_stat.h>
20#include <linux/swap.h>
21#include <linux/pagemap.h>
22#include <linux/init.h>
23#include <linux/highmem.h>
Anton Vorontsov70ddf632013-04-29 15:08:31 -070024#include <linux/vmpressure.h>
Andrew Mortone129b5c2006-09-27 01:50:00 -070025#include <linux/vmstat.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070026#include <linux/file.h>
27#include <linux/writeback.h>
28#include <linux/blkdev.h>
29#include <linux/buffer_head.h> /* for try_to_release_page(),
30 buffer_heads_over_limit */
31#include <linux/mm_inline.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070032#include <linux/backing-dev.h>
33#include <linux/rmap.h>
34#include <linux/topology.h>
35#include <linux/cpu.h>
36#include <linux/cpuset.h>
Mel Gorman3e7d3442011-01-13 15:45:56 -080037#include <linux/compaction.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070038#include <linux/notifier.h>
39#include <linux/rwsem.h>
Rafael J. Wysocki248a0302006-03-22 00:09:04 -080040#include <linux/delay.h>
Yasunori Goto3218ae12006-06-27 02:53:33 -070041#include <linux/kthread.h>
Nigel Cunningham7dfb7102006-12-06 20:34:23 -080042#include <linux/freezer.h>
Balbir Singh66e17072008-02-07 00:13:56 -080043#include <linux/memcontrol.h>
Keika Kobayashi873b4772008-07-25 01:48:52 -070044#include <linux/delayacct.h>
Lee Schermerhornaf936a12008-10-18 20:26:53 -070045#include <linux/sysctl.h>
KOSAKI Motohiro929bea72011-04-14 15:22:12 -070046#include <linux/oom.h>
Linus Torvalds268bb0c2011-05-20 12:50:29 -070047#include <linux/prefetch.h>
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -070048#include <linux/printk.h>
Ross Zwislerf9fe48b2016-01-22 15:10:40 -080049#include <linux/dax.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070050
51#include <asm/tlbflush.h>
52#include <asm/div64.h>
53
54#include <linux/swapops.h>
Rafael Aquini117aad12013-09-30 13:45:16 -070055#include <linux/balloon_compaction.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070056
Nick Piggin0f8053a2006-03-22 00:08:33 -080057#include "internal.h"
58
Mel Gorman33906bc2010-08-09 17:19:16 -070059#define CREATE_TRACE_POINTS
60#include <trace/events/vmscan.h>
61
Linus Torvalds1da177e2005-04-16 15:20:36 -070062struct scan_control {
KOSAKI Motohiro22fba332009-12-14 17:59:10 -080063 /* How many pages shrink_list() should reclaim */
64 unsigned long nr_to_reclaim;
65
Linus Torvalds1da177e2005-04-16 15:20:36 -070066 /* This context's GFP mask */
Al Viro6daa0e22005-10-21 03:18:50 -040067 gfp_t gfp_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -070068
Johannes Weineree814fe2014-08-06 16:06:19 -070069 /* Allocation order */
Andy Whitcroft5ad333e2007-07-17 04:03:16 -070070 int order;
Balbir Singh66e17072008-02-07 00:13:56 -080071
Johannes Weineree814fe2014-08-06 16:06:19 -070072 /*
73 * Nodemask of nodes allowed by the caller. If NULL, all nodes
74 * are scanned.
75 */
76 nodemask_t *nodemask;
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -070077
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -070078 /*
Johannes Weinerf16015f2012-01-12 17:17:52 -080079 * The memory cgroup that hit its limit and as a result is the
80 * primary target of this reclaim invocation.
81 */
82 struct mem_cgroup *target_mem_cgroup;
Balbir Singh66e17072008-02-07 00:13:56 -080083
Johannes Weineree814fe2014-08-06 16:06:19 -070084 /* Scan (total_size >> priority) pages at once */
85 int priority;
86
87 unsigned int may_writepage:1;
88
89 /* Can mapped pages be reclaimed? */
90 unsigned int may_unmap:1;
91
92 /* Can pages be swapped as part of reclaim? */
93 unsigned int may_swap:1;
94
Johannes Weiner241994ed2015-02-11 15:26:06 -080095 /* Can cgroups be reclaimed below their normal consumption range? */
96 unsigned int may_thrash:1;
97
Johannes Weineree814fe2014-08-06 16:06:19 -070098 unsigned int hibernation_mode:1;
99
100 /* One of the zones is ready for compaction */
101 unsigned int compaction_ready:1;
102
103 /* Incremented by the number of inactive pages that were scanned */
104 unsigned long nr_scanned;
105
106 /* Number of pages freed so far during a call to shrink_zones() */
107 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700108};
109
Linus Torvalds1da177e2005-04-16 15:20:36 -0700110#ifdef ARCH_HAS_PREFETCH
111#define prefetch_prev_lru_page(_page, _base, _field) \
112 do { \
113 if ((_page)->lru.prev != _base) { \
114 struct page *prev; \
115 \
116 prev = lru_to_page(&(_page->lru)); \
117 prefetch(&prev->_field); \
118 } \
119 } while (0)
120#else
121#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
122#endif
123
124#ifdef ARCH_HAS_PREFETCHW
125#define prefetchw_prev_lru_page(_page, _base, _field) \
126 do { \
127 if ((_page)->lru.prev != _base) { \
128 struct page *prev; \
129 \
130 prev = lru_to_page(&(_page->lru)); \
131 prefetchw(&prev->_field); \
132 } \
133 } while (0)
134#else
135#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
136#endif
137
138/*
139 * From 0 .. 100. Higher means more swappy.
140 */
141int vm_swappiness = 60;
Wang Sheng-Huid0480be2014-08-06 16:07:07 -0700142/*
143 * The total number of pages which are beyond the high watermark within all
144 * zones.
145 */
146unsigned long vm_total_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700147
148static LIST_HEAD(shrinker_list);
149static DECLARE_RWSEM(shrinker_rwsem);
150
Andrew Mortonc255a452012-07-31 16:43:02 -0700151#ifdef CONFIG_MEMCG
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800152static bool global_reclaim(struct scan_control *sc)
153{
Johannes Weinerf16015f2012-01-12 17:17:52 -0800154 return !sc->target_mem_cgroup;
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800155}
Tejun Heo97c93412015-05-22 18:23:36 -0400156
157/**
158 * sane_reclaim - is the usual dirty throttling mechanism operational?
159 * @sc: scan_control in question
160 *
161 * The normal page dirty throttling mechanism in balance_dirty_pages() is
162 * completely broken with the legacy memcg and direct stalling in
163 * shrink_page_list() is used for throttling instead, which lacks all the
164 * niceties such as fairness, adaptive pausing, bandwidth proportional
165 * allocation and configurability.
166 *
167 * This function tests whether the vmscan currently in progress can assume
168 * that the normal dirty throttling mechanism is operational.
169 */
170static bool sane_reclaim(struct scan_control *sc)
171{
172 struct mem_cgroup *memcg = sc->target_mem_cgroup;
173
174 if (!memcg)
175 return true;
176#ifdef CONFIG_CGROUP_WRITEBACK
Linus Torvalds69234ac2015-11-05 14:51:32 -0800177 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
Tejun Heo97c93412015-05-22 18:23:36 -0400178 return true;
179#endif
180 return false;
181}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800182#else
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800183static bool global_reclaim(struct scan_control *sc)
184{
185 return true;
186}
Tejun Heo97c93412015-05-22 18:23:36 -0400187
188static bool sane_reclaim(struct scan_control *sc)
189{
190 return true;
191}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800192#endif
193
Johannes Weinera1c3bfb2014-01-29 14:05:41 -0800194static unsigned long zone_reclaimable_pages(struct zone *zone)
Lisa Du6e543d52013-09-11 14:22:36 -0700195{
Alexandru Moised031a152015-11-05 18:48:08 -0800196 unsigned long nr;
Lisa Du6e543d52013-09-11 14:22:36 -0700197
Michal Hocko0db2cb82016-03-15 14:57:01 -0700198 nr = zone_page_state_snapshot(zone, NR_ACTIVE_FILE) +
199 zone_page_state_snapshot(zone, NR_INACTIVE_FILE) +
200 zone_page_state_snapshot(zone, NR_ISOLATED_FILE);
Lisa Du6e543d52013-09-11 14:22:36 -0700201
202 if (get_nr_swap_pages() > 0)
Michal Hocko0db2cb82016-03-15 14:57:01 -0700203 nr += zone_page_state_snapshot(zone, NR_ACTIVE_ANON) +
204 zone_page_state_snapshot(zone, NR_INACTIVE_ANON) +
205 zone_page_state_snapshot(zone, NR_ISOLATED_ANON);
Lisa Du6e543d52013-09-11 14:22:36 -0700206
207 return nr;
208}
209
210bool zone_reclaimable(struct zone *zone)
211{
Michal Hocko0db2cb82016-03-15 14:57:01 -0700212 return zone_page_state_snapshot(zone, NR_PAGES_SCANNED) <
Mel Gorman0d5d8232014-08-06 16:07:16 -0700213 zone_reclaimable_pages(zone) * 6;
Lisa Du6e543d52013-09-11 14:22:36 -0700214}
215
Johannes Weiner23047a92016-03-15 14:57:16 -0700216unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru)
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800217{
Hugh Dickinsc3c787e2012-05-29 15:06:52 -0700218 if (!mem_cgroup_disabled())
Hugh Dickins4d7dcca2012-05-29 15:07:08 -0700219 return mem_cgroup_get_lru_size(lruvec, lru);
KOSAKI Motohiroa3d8e052009-01-07 18:08:19 -0800220
Konstantin Khlebnikov074291f2012-05-29 15:07:00 -0700221 return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800222}
223
Linus Torvalds1da177e2005-04-16 15:20:36 -0700224/*
Glauber Costa1d3d4432013-08-28 10:18:04 +1000225 * Add a shrinker callback to be called from the vm.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700226 */
Glauber Costa1d3d4432013-08-28 10:18:04 +1000227int register_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700228{
Glauber Costa1d3d4432013-08-28 10:18:04 +1000229 size_t size = sizeof(*shrinker->nr_deferred);
230
Glauber Costa1d3d4432013-08-28 10:18:04 +1000231 if (shrinker->flags & SHRINKER_NUMA_AWARE)
232 size *= nr_node_ids;
233
234 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
235 if (!shrinker->nr_deferred)
236 return -ENOMEM;
237
Rusty Russell8e1f9362007-07-17 04:03:17 -0700238 down_write(&shrinker_rwsem);
239 list_add_tail(&shrinker->list, &shrinker_list);
240 up_write(&shrinker_rwsem);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000241 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700242}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700243EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700244
245/*
246 * Remove one
247 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700248void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700249{
250 down_write(&shrinker_rwsem);
251 list_del(&shrinker->list);
252 up_write(&shrinker_rwsem);
Andrew Vaginae393322013-10-16 13:46:46 -0700253 kfree(shrinker->nr_deferred);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700254}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700255EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700256
257#define SHRINK_BATCH 128
Glauber Costa1d3d4432013-08-28 10:18:04 +1000258
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800259static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
260 struct shrinker *shrinker,
261 unsigned long nr_scanned,
262 unsigned long nr_eligible)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263{
Glauber Costa1d3d4432013-08-28 10:18:04 +1000264 unsigned long freed = 0;
265 unsigned long long delta;
266 long total_scan;
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700267 long freeable;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000268 long nr;
269 long new_nr;
270 int nid = shrinkctl->nid;
271 long batch_size = shrinker->batch ? shrinker->batch
272 : SHRINK_BATCH;
273
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700274 freeable = shrinker->count_objects(shrinker, shrinkctl);
275 if (freeable == 0)
Glauber Costa1d3d4432013-08-28 10:18:04 +1000276 return 0;
277
278 /*
279 * copy the current shrinker scan count into a local variable
280 * and zero it so that other concurrent shrinker invocations
281 * don't also do this scanning work.
282 */
283 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
284
285 total_scan = nr;
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800286 delta = (4 * nr_scanned) / shrinker->seeks;
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700287 delta *= freeable;
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800288 do_div(delta, nr_eligible + 1);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000289 total_scan += delta;
290 if (total_scan < 0) {
Pintu Kumar8612c662014-12-10 15:42:58 -0800291 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
Dave Chinnera0b02132013-08-28 10:18:16 +1000292 shrinker->scan_objects, total_scan);
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700293 total_scan = freeable;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000294 }
295
296 /*
297 * We need to avoid excessive windup on filesystem shrinkers
298 * due to large numbers of GFP_NOFS allocations causing the
299 * shrinkers to return -1 all the time. This results in a large
300 * nr being built up so when a shrink that can do some work
301 * comes along it empties the entire cache due to nr >>>
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700302 * freeable. This is bad for sustaining a working set in
Glauber Costa1d3d4432013-08-28 10:18:04 +1000303 * memory.
304 *
305 * Hence only allow the shrinker to scan the entire cache when
306 * a large delta change is calculated directly.
307 */
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700308 if (delta < freeable / 4)
309 total_scan = min(total_scan, freeable / 2);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000310
311 /*
312 * Avoid risking looping forever due to too large nr value:
313 * never try to free more than twice the estimate number of
314 * freeable entries.
315 */
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700316 if (total_scan > freeable * 2)
317 total_scan = freeable * 2;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000318
319 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800320 nr_scanned, nr_eligible,
321 freeable, delta, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000322
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800323 /*
324 * Normally, we should not scan less than batch_size objects in one
325 * pass to avoid too frequent shrinker calls, but if the slab has less
326 * than batch_size objects in total and we are really tight on memory,
327 * we will try to reclaim all available objects, otherwise we can end
328 * up failing allocations although there are plenty of reclaimable
329 * objects spread over several slabs with usage less than the
330 * batch_size.
331 *
332 * We detect the "tight on memory" situations by looking at the total
333 * number of objects we want to scan (total_scan). If it is greater
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700334 * than the total number of objects on slab (freeable), we must be
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800335 * scanning at high prio and therefore should try to reclaim as much as
336 * possible.
337 */
338 while (total_scan >= batch_size ||
Vladimir Davydovd5bc5fd2014-04-03 14:47:32 -0700339 total_scan >= freeable) {
Dave Chinnera0b02132013-08-28 10:18:16 +1000340 unsigned long ret;
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800341 unsigned long nr_to_scan = min(batch_size, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000342
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800343 shrinkctl->nr_to_scan = nr_to_scan;
Dave Chinnera0b02132013-08-28 10:18:16 +1000344 ret = shrinker->scan_objects(shrinker, shrinkctl);
345 if (ret == SHRINK_STOP)
346 break;
347 freed += ret;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000348
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800349 count_vm_events(SLABS_SCANNED, nr_to_scan);
350 total_scan -= nr_to_scan;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000351
352 cond_resched();
353 }
354
355 /*
356 * move the unused scan count back into the shrinker in a
357 * manner that handles concurrent updates. If we exhausted the
358 * scan, there is no need to do an update.
359 */
360 if (total_scan > 0)
361 new_nr = atomic_long_add_return(total_scan,
362 &shrinker->nr_deferred[nid]);
363 else
364 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
365
Dave Hansendf9024a2014-06-04 16:08:07 -0700366 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000367 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700368}
369
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800370/**
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800371 * shrink_slab - shrink slab caches
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800372 * @gfp_mask: allocation context
373 * @nid: node whose slab caches to target
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800374 * @memcg: memory cgroup whose slab caches to target
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800375 * @nr_scanned: pressure numerator
376 * @nr_eligible: pressure denominator
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800378 * Call the shrink functions to age shrinkable caches.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700379 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800380 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
381 * unaware shrinkers will receive a node id of 0 instead.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700382 *
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800383 * @memcg specifies the memory cgroup to target. If it is not NULL,
384 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
Vladimir Davydov0fc9f582016-03-17 14:18:30 -0700385 * objects from the memory cgroup specified. Otherwise, only unaware
386 * shrinkers are called.
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800387 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800388 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
389 * the available objects should be scanned. Page reclaim for example
390 * passes the number of pages scanned and the number of pages on the
391 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
392 * when it encountered mapped pages. The ratio is further biased by
393 * the ->seeks setting of the shrink function, which indicates the
394 * cost to recreate an object relative to that of an LRU page.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700395 *
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800396 * Returns the number of reclaimed slab objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700397 */
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800398static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
399 struct mem_cgroup *memcg,
400 unsigned long nr_scanned,
401 unsigned long nr_eligible)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700402{
403 struct shrinker *shrinker;
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000404 unsigned long freed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700405
Vladimir Davydov0fc9f582016-03-17 14:18:30 -0700406 if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg)))
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800407 return 0;
408
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800409 if (nr_scanned == 0)
410 nr_scanned = SWAP_CLUSTER_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411
Minchan Kimf06590b2011-05-24 17:11:11 -0700412 if (!down_read_trylock(&shrinker_rwsem)) {
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000413 /*
414 * If we would return 0, our callers would understand that we
415 * have nothing else to shrink and give up trying. By returning
416 * 1 we keep it going and assume we'll be able to shrink next
417 * time.
418 */
419 freed = 1;
Minchan Kimf06590b2011-05-24 17:11:11 -0700420 goto out;
421 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422
423 list_for_each_entry(shrinker, &shrinker_list, list) {
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800424 struct shrink_control sc = {
425 .gfp_mask = gfp_mask,
426 .nid = nid,
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800427 .memcg = memcg,
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800428 };
Vladimir Davydovec970972014-01-23 15:53:23 -0800429
Vladimir Davydov0fc9f582016-03-17 14:18:30 -0700430 /*
431 * If kernel memory accounting is disabled, we ignore
432 * SHRINKER_MEMCG_AWARE flag and call all shrinkers
433 * passing NULL for memcg.
434 */
435 if (memcg_kmem_enabled() &&
436 !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE))
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800437 continue;
438
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800439 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
440 sc.nid = 0;
Dave Chinneracf92b42011-07-08 14:14:35 +1000441
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800442 freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443 }
Johannes Weiner6b4f7792014-12-12 16:56:13 -0800444
Linus Torvalds1da177e2005-04-16 15:20:36 -0700445 up_read(&shrinker_rwsem);
Minchan Kimf06590b2011-05-24 17:11:11 -0700446out:
447 cond_resched();
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000448 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700449}
450
Vladimir Davydovcb731d62015-02-12 14:58:54 -0800451void drop_slab_node(int nid)
452{
453 unsigned long freed;
454
455 do {
456 struct mem_cgroup *memcg = NULL;
457
458 freed = 0;
459 do {
460 freed += shrink_slab(GFP_KERNEL, nid, memcg,
461 1000, 1000);
462 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
463 } while (freed > 10);
464}
465
466void drop_slab(void)
467{
468 int nid;
469
470 for_each_online_node(nid)
471 drop_slab_node(nid);
472}
473
Linus Torvalds1da177e2005-04-16 15:20:36 -0700474static inline int is_page_cache_freeable(struct page *page)
475{
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700476 /*
477 * A freeable page cache page is referenced only by the caller
478 * that isolated the page, the page cache radix tree and
479 * optional buffer heads at page->private.
480 */
Johannes Weineredcf4742009-09-21 17:02:59 -0700481 return page_count(page) - page_has_private(page) == 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700482}
483
Tejun Heo703c2702015-05-22 17:13:44 -0400484static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700485{
Christoph Lameter930d9152006-01-08 01:00:47 -0800486 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700487 return 1;
Tejun Heo703c2702015-05-22 17:13:44 -0400488 if (!inode_write_congested(inode))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700489 return 1;
Tejun Heo703c2702015-05-22 17:13:44 -0400490 if (inode_to_bdi(inode) == current->backing_dev_info)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700491 return 1;
492 return 0;
493}
494
495/*
496 * We detected a synchronous write error writing a page out. Probably
497 * -ENOSPC. We need to propagate that into the address_space for a subsequent
498 * fsync(), msync() or close().
499 *
500 * The tricky part is that after writepage we cannot touch the mapping: nothing
501 * prevents it from being freed up. But we have a ref on the page and once
502 * that page is locked, the mapping is pinned.
503 *
504 * We're allowed to run sleeping lock_page() here because we know the caller has
505 * __GFP_FS.
506 */
507static void handle_write_error(struct address_space *mapping,
508 struct page *page, int error)
509{
Jens Axboe7eaceac2011-03-10 08:52:07 +0100510 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700511 if (page_mapping(page) == mapping)
512 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700513 unlock_page(page);
514}
515
Christoph Lameter04e62a22006-06-23 02:03:38 -0700516/* possible outcome of pageout() */
517typedef enum {
518 /* failed to write page out, page is locked */
519 PAGE_KEEP,
520 /* move page to the active list, page is locked */
521 PAGE_ACTIVATE,
522 /* page has been sent to the disk successfully, page is unlocked */
523 PAGE_SUCCESS,
524 /* page is clean and locked */
525 PAGE_CLEAN,
526} pageout_t;
527
Linus Torvalds1da177e2005-04-16 15:20:36 -0700528/*
Andrew Morton1742f192006-03-22 00:08:21 -0800529 * pageout is called by shrink_page_list() for each dirty page.
530 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700531 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700532static pageout_t pageout(struct page *page, struct address_space *mapping,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700533 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700534{
535 /*
536 * If the page is dirty, only perform writeback if that write
537 * will be non-blocking. To prevent this allocation from being
538 * stalled by pagecache activity. But note that there may be
539 * stalls if we need to run get_block(). We could test
540 * PagePrivate for that.
541 *
Al Viro81742022014-04-03 03:17:43 -0400542 * If this process is currently in __generic_file_write_iter() against
Linus Torvalds1da177e2005-04-16 15:20:36 -0700543 * this page's queue, we can perform writeback even if that
544 * will block.
545 *
546 * If the page is swapcache, write it back even if that would
547 * block, for some throttling. This happens by accident, because
548 * swap_backing_dev_info is bust: it doesn't reflect the
549 * congestion state of the swapdevs. Easy to fix, if needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700550 */
551 if (!is_page_cache_freeable(page))
552 return PAGE_KEEP;
553 if (!mapping) {
554 /*
555 * Some data journaling orphaned pages can have
556 * page->mapping == NULL while being dirty with clean buffers.
557 */
David Howells266cf652009-04-03 16:42:36 +0100558 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700559 if (try_to_free_buffers(page)) {
560 ClearPageDirty(page);
Mitchel Humpherysb1de0d12014-06-06 14:38:30 -0700561 pr_info("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700562 return PAGE_CLEAN;
563 }
564 }
565 return PAGE_KEEP;
566 }
567 if (mapping->a_ops->writepage == NULL)
568 return PAGE_ACTIVATE;
Tejun Heo703c2702015-05-22 17:13:44 -0400569 if (!may_write_to_inode(mapping->host, sc))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700570 return PAGE_KEEP;
571
572 if (clear_page_dirty_for_io(page)) {
573 int res;
574 struct writeback_control wbc = {
575 .sync_mode = WB_SYNC_NONE,
576 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700577 .range_start = 0,
578 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700579 .for_reclaim = 1,
580 };
581
582 SetPageReclaim(page);
583 res = mapping->a_ops->writepage(page, &wbc);
584 if (res < 0)
585 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800586 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700587 ClearPageReclaim(page);
588 return PAGE_ACTIVATE;
589 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700590
Linus Torvalds1da177e2005-04-16 15:20:36 -0700591 if (!PageWriteback(page)) {
592 /* synchronous write or broken a_ops? */
593 ClearPageReclaim(page);
594 }
yalin wang3aa23852016-01-14 15:18:30 -0800595 trace_mm_vmscan_writepage(page);
Andrew Mortone129b5c2006-09-27 01:50:00 -0700596 inc_zone_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700597 return PAGE_SUCCESS;
598 }
599
600 return PAGE_CLEAN;
601}
602
Andrew Mortona649fd92006-10-17 00:09:36 -0700603/*
Nick Piggine2867812008-07-25 19:45:30 -0700604 * Same as remove_mapping, but if the page is removed from the mapping, it
605 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700606 */
Johannes Weinera5289102014-04-03 14:47:51 -0700607static int __remove_mapping(struct address_space *mapping, struct page *page,
608 bool reclaimed)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800609{
Greg Thelenc4843a72015-05-22 17:13:16 -0400610 unsigned long flags;
Greg Thelenc4843a72015-05-22 17:13:16 -0400611
Nick Piggin28e4d962006-09-25 23:31:23 -0700612 BUG_ON(!PageLocked(page));
613 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800614
Greg Thelenc4843a72015-05-22 17:13:16 -0400615 spin_lock_irqsave(&mapping->tree_lock, flags);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800616 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700617 * The non racy check for a busy page.
618 *
619 * Must be careful with the order of the tests. When someone has
620 * a ref to the page, it may be possible that they dirty it then
621 * drop the reference. So if PageDirty is tested before page_count
622 * here, then the following race may occur:
623 *
624 * get_user_pages(&page);
625 * [user mapping goes away]
626 * write_to(page);
627 * !PageDirty(page) [good]
628 * SetPageDirty(page);
629 * put_page(page);
630 * !page_count(page) [good, discard it]
631 *
632 * [oops, our write_to data is lost]
633 *
634 * Reversing the order of the tests ensures such a situation cannot
635 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
636 * load is not satisfied before that of page->_count.
637 *
638 * Note that if SetPageDirty is always performed via set_page_dirty,
639 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800640 */
Joonsoo Kimfe896d12016-03-17 14:19:26 -0700641 if (!page_ref_freeze(page, 2))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800642 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700643 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
644 if (unlikely(PageDirty(page))) {
Joonsoo Kimfe896d12016-03-17 14:19:26 -0700645 page_ref_unfreeze(page, 2);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800646 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700647 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800648
649 if (PageSwapCache(page)) {
650 swp_entry_t swap = { .val = page_private(page) };
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700651 mem_cgroup_swapout(page, swap);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800652 __delete_from_swap_cache(page);
Greg Thelenc4843a72015-05-22 17:13:16 -0400653 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Johannes Weiner0a31bc92014-08-08 14:19:22 -0700654 swapcache_free(swap);
Nick Piggine2867812008-07-25 19:45:30 -0700655 } else {
Linus Torvalds6072d132010-12-01 13:35:19 -0500656 void (*freepage)(struct page *);
Johannes Weinera5289102014-04-03 14:47:51 -0700657 void *shadow = NULL;
Linus Torvalds6072d132010-12-01 13:35:19 -0500658
659 freepage = mapping->a_ops->freepage;
Johannes Weinera5289102014-04-03 14:47:51 -0700660 /*
661 * Remember a shadow entry for reclaimed file cache in
662 * order to detect refaults, thus thrashing, later on.
663 *
664 * But don't store shadows in an address space that is
665 * already exiting. This is not just an optizimation,
666 * inode reclaim needs to empty out the radix tree or
667 * the nodes are lost. Don't plant shadows behind its
668 * back.
Ross Zwislerf9fe48b2016-01-22 15:10:40 -0800669 *
670 * We also don't store shadows for DAX mappings because the
671 * only page cache pages found in these are zero pages
672 * covering holes, and because we don't want to mix DAX
673 * exceptional entries and shadow exceptional entries in the
674 * same page_tree.
Johannes Weinera5289102014-04-03 14:47:51 -0700675 */
676 if (reclaimed && page_is_file_cache(page) &&
Ross Zwislerf9fe48b2016-01-22 15:10:40 -0800677 !mapping_exiting(mapping) && !dax_mapping(mapping))
Johannes Weinera5289102014-04-03 14:47:51 -0700678 shadow = workingset_eviction(mapping, page);
Johannes Weiner62cccb82016-03-15 14:57:22 -0700679 __delete_from_page_cache(page, shadow);
Greg Thelenc4843a72015-05-22 17:13:16 -0400680 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Linus Torvalds6072d132010-12-01 13:35:19 -0500681
682 if (freepage != NULL)
683 freepage(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800684 }
685
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800686 return 1;
687
688cannot_free:
Greg Thelenc4843a72015-05-22 17:13:16 -0400689 spin_unlock_irqrestore(&mapping->tree_lock, flags);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800690 return 0;
691}
692
Linus Torvalds1da177e2005-04-16 15:20:36 -0700693/*
Nick Piggine2867812008-07-25 19:45:30 -0700694 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
695 * someone else has a ref on the page, abort and return 0. If it was
696 * successfully detached, return 1. Assumes the caller has a single ref on
697 * this page.
698 */
699int remove_mapping(struct address_space *mapping, struct page *page)
700{
Johannes Weinera5289102014-04-03 14:47:51 -0700701 if (__remove_mapping(mapping, page, false)) {
Nick Piggine2867812008-07-25 19:45:30 -0700702 /*
703 * Unfreezing the refcount with 1 rather than 2 effectively
704 * drops the pagecache ref for us without requiring another
705 * atomic operation.
706 */
Joonsoo Kimfe896d12016-03-17 14:19:26 -0700707 page_ref_unfreeze(page, 1);
Nick Piggine2867812008-07-25 19:45:30 -0700708 return 1;
709 }
710 return 0;
711}
712
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700713/**
714 * putback_lru_page - put previously isolated page onto appropriate LRU list
715 * @page: page to be put back to appropriate lru list
716 *
717 * Add previously isolated @page to appropriate LRU list.
718 * Page may still be unevictable for other reasons.
719 *
720 * lru_lock must not be held, interrupts must be enabled.
721 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700722void putback_lru_page(struct page *page)
723{
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700724 bool is_unevictable;
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700725 int was_unevictable = PageUnevictable(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700726
Sasha Levin309381fea2014-01-23 15:52:54 -0800727 VM_BUG_ON_PAGE(PageLRU(page), page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700728
729redo:
730 ClearPageUnevictable(page);
731
Hugh Dickins39b5f292012-10-08 16:33:18 -0700732 if (page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700733 /*
734 * For evictable pages, we can use the cache.
735 * In event of a race, worst case is we end up with an
736 * unevictable page on [in]active list.
737 * We know how to handle that.
738 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700739 is_unevictable = false;
Mel Gormanc53954a2013-07-03 15:02:34 -0700740 lru_cache_add(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700741 } else {
742 /*
743 * Put unevictable pages directly on zone's unevictable
744 * list.
745 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700746 is_unevictable = true;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700747 add_page_to_unevictable_list(page);
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700748 /*
Minchan Kim21ee9f32011-10-31 17:09:28 -0700749 * When racing with an mlock or AS_UNEVICTABLE clearing
750 * (page is unlocked) make sure that if the other thread
751 * does not observe our setting of PG_lru and fails
Hugh Dickins24513262012-01-20 14:34:21 -0800752 * isolation/check_move_unevictable_pages,
Minchan Kim21ee9f32011-10-31 17:09:28 -0700753 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700754 * the page back to the evictable list.
755 *
Minchan Kim21ee9f32011-10-31 17:09:28 -0700756 * The other side is TestClearPageMlocked() or shmem_lock().
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700757 */
758 smp_mb();
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700759 }
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700760
761 /*
762 * page's status can change while we move it among lru. If an evictable
763 * page is on unevictable list, it never be freed. To avoid that,
764 * check after we added it to the list, again.
765 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700766 if (is_unevictable && page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700767 if (!isolate_lru_page(page)) {
768 put_page(page);
769 goto redo;
770 }
771 /* This means someone else dropped this page from LRU
772 * So, it will be freed or putback to LRU again. There is
773 * nothing to do here.
774 */
775 }
776
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700777 if (was_unevictable && !is_unevictable)
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700778 count_vm_event(UNEVICTABLE_PGRESCUED);
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700779 else if (!was_unevictable && is_unevictable)
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700780 count_vm_event(UNEVICTABLE_PGCULLED);
781
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700782 put_page(page); /* drop ref from isolate */
783}
784
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800785enum page_references {
786 PAGEREF_RECLAIM,
787 PAGEREF_RECLAIM_CLEAN,
Johannes Weiner645747462010-03-05 13:42:22 -0800788 PAGEREF_KEEP,
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800789 PAGEREF_ACTIVATE,
790};
791
792static enum page_references page_check_references(struct page *page,
793 struct scan_control *sc)
794{
Johannes Weiner645747462010-03-05 13:42:22 -0800795 int referenced_ptes, referenced_page;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800796 unsigned long vm_flags;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800797
Johannes Weinerc3ac9a82012-05-29 15:06:25 -0700798 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
799 &vm_flags);
Johannes Weiner645747462010-03-05 13:42:22 -0800800 referenced_page = TestClearPageReferenced(page);
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800801
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800802 /*
803 * Mlock lost the isolation race with us. Let try_to_unmap()
804 * move the page to the unevictable list.
805 */
806 if (vm_flags & VM_LOCKED)
807 return PAGEREF_RECLAIM;
808
Johannes Weiner645747462010-03-05 13:42:22 -0800809 if (referenced_ptes) {
Michal Hockoe48982732012-05-29 15:06:45 -0700810 if (PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800811 return PAGEREF_ACTIVATE;
812 /*
813 * All mapped pages start out with page table
814 * references from the instantiating fault, so we need
815 * to look twice if a mapped file page is used more
816 * than once.
817 *
818 * Mark it and spare it for another trip around the
819 * inactive list. Another page table reference will
820 * lead to its activation.
821 *
822 * Note: the mark is set for activated pages as well
823 * so that recently deactivated but used pages are
824 * quickly recovered.
825 */
826 SetPageReferenced(page);
827
Konstantin Khlebnikov34dbc672012-01-10 15:06:59 -0800828 if (referenced_page || referenced_ptes > 1)
Johannes Weiner645747462010-03-05 13:42:22 -0800829 return PAGEREF_ACTIVATE;
830
Konstantin Khlebnikovc909e992012-01-10 15:07:03 -0800831 /*
832 * Activate file-backed executable pages after first usage.
833 */
834 if (vm_flags & VM_EXEC)
835 return PAGEREF_ACTIVATE;
836
Johannes Weiner645747462010-03-05 13:42:22 -0800837 return PAGEREF_KEEP;
838 }
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800839
840 /* Reclaim if clean, defer dirty pages to writeback */
KOSAKI Motohiro2e302442010-10-26 14:21:46 -0700841 if (referenced_page && !PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800842 return PAGEREF_RECLAIM_CLEAN;
843
844 return PAGEREF_RECLAIM;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800845}
846
Mel Gormane2be15f2013-07-03 15:01:57 -0700847/* Check if a page is dirty or under writeback */
848static void page_check_dirty_writeback(struct page *page,
849 bool *dirty, bool *writeback)
850{
Mel Gormanb4597222013-07-03 15:02:05 -0700851 struct address_space *mapping;
852
Mel Gormane2be15f2013-07-03 15:01:57 -0700853 /*
854 * Anonymous pages are not handled by flushers and must be written
855 * from reclaim context. Do not stall reclaim based on them
856 */
857 if (!page_is_file_cache(page)) {
858 *dirty = false;
859 *writeback = false;
860 return;
861 }
862
863 /* By default assume that the page flags are accurate */
864 *dirty = PageDirty(page);
865 *writeback = PageWriteback(page);
Mel Gormanb4597222013-07-03 15:02:05 -0700866
867 /* Verify dirty/writeback state if the filesystem supports it */
868 if (!page_has_private(page))
869 return;
870
871 mapping = page_mapping(page);
872 if (mapping && mapping->a_ops->is_dirty_writeback)
873 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
Mel Gormane2be15f2013-07-03 15:01:57 -0700874}
875
Nick Piggine2867812008-07-25 19:45:30 -0700876/*
Andrew Morton1742f192006-03-22 00:08:21 -0800877 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700878 */
Andrew Morton1742f192006-03-22 00:08:21 -0800879static unsigned long shrink_page_list(struct list_head *page_list,
Konstantin Khlebnikov6a18adb2012-05-29 15:06:59 -0700880 struct zone *zone,
Mel Gormanf84f6e22011-10-31 17:07:51 -0700881 struct scan_control *sc,
Minchan Kim02c6de82012-10-08 16:31:55 -0700882 enum ttu_flags ttu_flags,
Mel Gorman8e950282013-07-03 15:02:02 -0700883 unsigned long *ret_nr_dirty,
Mel Gormand43006d2013-07-03 15:01:50 -0700884 unsigned long *ret_nr_unqueued_dirty,
Mel Gorman8e950282013-07-03 15:02:02 -0700885 unsigned long *ret_nr_congested,
Minchan Kim02c6de82012-10-08 16:31:55 -0700886 unsigned long *ret_nr_writeback,
Mel Gormanb1a6f212013-07-03 15:01:58 -0700887 unsigned long *ret_nr_immediate,
Minchan Kim02c6de82012-10-08 16:31:55 -0700888 bool force_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889{
890 LIST_HEAD(ret_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700891 LIST_HEAD(free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700892 int pgactivate = 0;
Mel Gormand43006d2013-07-03 15:01:50 -0700893 unsigned long nr_unqueued_dirty = 0;
Mel Gorman0e093d992010-10-26 14:21:45 -0700894 unsigned long nr_dirty = 0;
895 unsigned long nr_congested = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800896 unsigned long nr_reclaimed = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -0700897 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -0700898 unsigned long nr_immediate = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700899
900 cond_resched();
901
Linus Torvalds1da177e2005-04-16 15:20:36 -0700902 while (!list_empty(page_list)) {
903 struct address_space *mapping;
904 struct page *page;
905 int may_enter_fs;
Minchan Kim02c6de82012-10-08 16:31:55 -0700906 enum page_references references = PAGEREF_RECLAIM_CLEAN;
Mel Gormane2be15f2013-07-03 15:01:57 -0700907 bool dirty, writeback;
Minchan Kim854e9ed2016-01-15 16:54:53 -0800908 bool lazyfree = false;
909 int ret = SWAP_SUCCESS;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700910
911 cond_resched();
912
913 page = lru_to_page(page_list);
914 list_del(&page->lru);
915
Nick Piggin529ae9a2008-08-02 12:01:03 +0200916 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700917 goto keep;
918
Sasha Levin309381fea2014-01-23 15:52:54 -0800919 VM_BUG_ON_PAGE(PageActive(page), page);
920 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700921
922 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800923
Hugh Dickins39b5f292012-10-08 16:33:18 -0700924 if (unlikely(!page_evictable(page)))
Nick Pigginb291f002008-10-18 20:26:44 -0700925 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700926
Johannes Weinera6dc60f82009-03-31 15:19:30 -0700927 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800928 goto keep_locked;
929
Linus Torvalds1da177e2005-04-16 15:20:36 -0700930 /* Double the slab pressure for mapped and swapcache pages */
931 if (page_mapped(page) || PageSwapCache(page))
932 sc->nr_scanned++;
933
Andy Whitcroftc661b072007-08-22 14:01:26 -0700934 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
935 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
936
Mel Gorman283aba92013-07-03 15:01:51 -0700937 /*
Mel Gormane2be15f2013-07-03 15:01:57 -0700938 * The number of dirty pages determines if a zone is marked
939 * reclaim_congested which affects wait_iff_congested. kswapd
940 * will stall and start writing pages if the tail of the LRU
941 * is all dirty unqueued pages.
942 */
943 page_check_dirty_writeback(page, &dirty, &writeback);
944 if (dirty || writeback)
945 nr_dirty++;
946
947 if (dirty && !writeback)
948 nr_unqueued_dirty++;
949
Mel Gormand04e8ac2013-07-03 15:02:03 -0700950 /*
951 * Treat this page as congested if the underlying BDI is or if
952 * pages are cycling through the LRU so quickly that the
953 * pages marked for immediate reclaim are making it to the
954 * end of the LRU a second time.
955 */
Mel Gormane2be15f2013-07-03 15:01:57 -0700956 mapping = page_mapping(page);
Jamie Liu1da58ee2014-12-10 15:43:20 -0800957 if (((dirty || writeback) && mapping &&
Tejun Heo703c2702015-05-22 17:13:44 -0400958 inode_write_congested(mapping->host)) ||
Mel Gormand04e8ac2013-07-03 15:02:03 -0700959 (writeback && PageReclaim(page)))
Mel Gormane2be15f2013-07-03 15:01:57 -0700960 nr_congested++;
961
962 /*
Mel Gorman283aba92013-07-03 15:01:51 -0700963 * If a page at the tail of the LRU is under writeback, there
964 * are three cases to consider.
965 *
966 * 1) If reclaim is encountering an excessive number of pages
967 * under writeback and this page is both under writeback and
968 * PageReclaim then it indicates that pages are being queued
969 * for IO but are being recycled through the LRU before the
970 * IO can complete. Waiting on the page itself risks an
971 * indefinite stall if it is impossible to writeback the
972 * page due to IO error or disconnected storage so instead
Mel Gormanb1a6f212013-07-03 15:01:58 -0700973 * note that the LRU is being scanned too quickly and the
974 * caller can stall after page list has been processed.
Mel Gorman283aba92013-07-03 15:01:51 -0700975 *
Tejun Heo97c93412015-05-22 18:23:36 -0400976 * 2) Global or new memcg reclaim encounters a page that is
Michal Hockoecf5fc62015-08-04 14:36:58 -0700977 * not marked for immediate reclaim, or the caller does not
978 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
979 * not to fs). In this case mark the page for immediate
Tejun Heo97c93412015-05-22 18:23:36 -0400980 * reclaim and continue scanning.
Mel Gorman283aba92013-07-03 15:01:51 -0700981 *
Michal Hockoecf5fc62015-08-04 14:36:58 -0700982 * Require may_enter_fs because we would wait on fs, which
983 * may not have submitted IO yet. And the loop driver might
Mel Gorman283aba92013-07-03 15:01:51 -0700984 * enter reclaim, and deadlock if it waits on a page for
985 * which it is needed to do the write (loop masks off
986 * __GFP_IO|__GFP_FS for this reason); but more thought
987 * would probably show more reasons.
988 *
Hugh Dickins7fadc822015-09-08 15:03:46 -0700989 * 3) Legacy memcg encounters a page that is already marked
Mel Gorman283aba92013-07-03 15:01:51 -0700990 * PageReclaim. memcg does not have any dirty pages
991 * throttling so we could easily OOM just because too many
992 * pages are in writeback and there is nothing else to
993 * reclaim. Wait for the writeback to complete.
994 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700995 if (PageWriteback(page)) {
Mel Gorman283aba92013-07-03 15:01:51 -0700996 /* Case 1 above */
997 if (current_is_kswapd() &&
998 PageReclaim(page) &&
Johannes Weiner57054652014-10-09 15:28:17 -0700999 test_bit(ZONE_WRITEBACK, &zone->flags)) {
Mel Gormanb1a6f212013-07-03 15:01:58 -07001000 nr_immediate++;
1001 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -07001002
1003 /* Case 2 above */
Tejun Heo97c93412015-05-22 18:23:36 -04001004 } else if (sane_reclaim(sc) ||
Michal Hockoecf5fc62015-08-04 14:36:58 -07001005 !PageReclaim(page) || !may_enter_fs) {
Hugh Dickinsc3b94f42012-07-31 16:45:59 -07001006 /*
1007 * This is slightly racy - end_page_writeback()
1008 * might have just cleared PageReclaim, then
1009 * setting PageReclaim here end up interpreted
1010 * as PageReadahead - but that does not matter
1011 * enough to care. What we do want is for this
1012 * page to have PageReclaim set next time memcg
1013 * reclaim reaches the tests above, so it will
1014 * then wait_on_page_writeback() to avoid OOM;
1015 * and it's also appropriate in global reclaim.
1016 */
1017 SetPageReclaim(page);
Michal Hockoe62e3842012-07-31 16:45:55 -07001018 nr_writeback++;
Hugh Dickinsc3b94f42012-07-31 16:45:59 -07001019 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -07001020
1021 /* Case 3 above */
1022 } else {
Hugh Dickins7fadc822015-09-08 15:03:46 -07001023 unlock_page(page);
Mel Gorman283aba92013-07-03 15:01:51 -07001024 wait_on_page_writeback(page);
Hugh Dickins7fadc822015-09-08 15:03:46 -07001025 /* then go back and try same page again */
1026 list_add_tail(&page->lru, page_list);
1027 continue;
Michal Hockoe62e3842012-07-31 16:45:55 -07001028 }
Andy Whitcroftc661b072007-08-22 14:01:26 -07001029 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001030
Minchan Kim02c6de82012-10-08 16:31:55 -07001031 if (!force_reclaim)
1032 references = page_check_references(page, sc);
1033
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001034 switch (references) {
1035 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001036 goto activate_locked;
Johannes Weiner645747462010-03-05 13:42:22 -08001037 case PAGEREF_KEEP:
1038 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001039 case PAGEREF_RECLAIM:
1040 case PAGEREF_RECLAIM_CLEAN:
1041 ; /* try to reclaim the page below */
1042 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001043
Linus Torvalds1da177e2005-04-16 15:20:36 -07001044 /*
1045 * Anonymous process memory has backing store?
1046 * Try to allocate it some swap space here.
1047 */
Nick Pigginb291f002008-10-18 20:26:44 -07001048 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -08001049 if (!(sc->gfp_mask & __GFP_IO))
1050 goto keep_locked;
Shaohua Li5bc7b8a2013-04-29 15:08:36 -07001051 if (!add_to_swap(page, page_list))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001052 goto activate_locked;
Minchan Kim854e9ed2016-01-15 16:54:53 -08001053 lazyfree = true;
Hugh Dickins63eb6b92008-11-19 15:36:37 -08001054 may_enter_fs = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001055
Mel Gormane2be15f2013-07-03 15:01:57 -07001056 /* Adding to swap updated mapping */
1057 mapping = page_mapping(page);
1058 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001059
1060 /*
1061 * The page is mapped into the page tables of one or more
1062 * processes. Try to unmap it here.
1063 */
1064 if (page_mapped(page) && mapping) {
Minchan Kim854e9ed2016-01-15 16:54:53 -08001065 switch (ret = try_to_unmap(page, lazyfree ?
1066 (ttu_flags | TTU_BATCH_FLUSH | TTU_LZFREE) :
1067 (ttu_flags | TTU_BATCH_FLUSH))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001068 case SWAP_FAIL:
1069 goto activate_locked;
1070 case SWAP_AGAIN:
1071 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -07001072 case SWAP_MLOCK:
1073 goto cull_mlocked;
Minchan Kim854e9ed2016-01-15 16:54:53 -08001074 case SWAP_LZFREE:
1075 goto lazyfree;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001076 case SWAP_SUCCESS:
1077 ; /* try to free the page below */
1078 }
1079 }
1080
1081 if (PageDirty(page)) {
Mel Gormanee728862011-10-31 17:07:38 -07001082 /*
1083 * Only kswapd can writeback filesystem pages to
Mel Gormand43006d2013-07-03 15:01:50 -07001084 * avoid risk of stack overflow but only writeback
1085 * if many dirty pages have been encountered.
Mel Gormanee728862011-10-31 17:07:38 -07001086 */
Mel Gormanf84f6e22011-10-31 17:07:51 -07001087 if (page_is_file_cache(page) &&
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001088 (!current_is_kswapd() ||
Johannes Weiner57054652014-10-09 15:28:17 -07001089 !test_bit(ZONE_DIRTY, &zone->flags))) {
Mel Gorman49ea7eb2011-10-31 17:07:59 -07001090 /*
1091 * Immediately reclaim when written back.
1092 * Similar in principal to deactivate_page()
1093 * except we already have the page isolated
1094 * and know it's dirty
1095 */
1096 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
1097 SetPageReclaim(page);
1098
Mel Gormanee728862011-10-31 17:07:38 -07001099 goto keep_locked;
1100 }
1101
Johannes Weinerdfc8d632010-03-05 13:42:19 -08001102 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001103 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -07001104 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001105 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -08001106 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001107 goto keep_locked;
1108
Mel Gormand950c942015-09-04 15:47:35 -07001109 /*
1110 * Page is dirty. Flush the TLB if a writable entry
1111 * potentially exists to avoid CPU writes after IO
1112 * starts and then write it out here.
1113 */
1114 try_to_unmap_flush_dirty();
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001115 switch (pageout(page, mapping, sc)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001116 case PAGE_KEEP:
1117 goto keep_locked;
1118 case PAGE_ACTIVATE:
1119 goto activate_locked;
1120 case PAGE_SUCCESS:
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001121 if (PageWriteback(page))
Mel Gorman41ac1992012-05-29 15:06:19 -07001122 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001123 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001124 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001125
Linus Torvalds1da177e2005-04-16 15:20:36 -07001126 /*
1127 * A synchronous write - probably a ramdisk. Go
1128 * ahead and try to reclaim the page.
1129 */
Nick Piggin529ae9a2008-08-02 12:01:03 +02001130 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001131 goto keep;
1132 if (PageDirty(page) || PageWriteback(page))
1133 goto keep_locked;
1134 mapping = page_mapping(page);
1135 case PAGE_CLEAN:
1136 ; /* try to free the page below */
1137 }
1138 }
1139
1140 /*
1141 * If the page has buffers, try to free the buffer mappings
1142 * associated with this page. If we succeed we try to free
1143 * the page as well.
1144 *
1145 * We do this even if the page is PageDirty().
1146 * try_to_release_page() does not perform I/O, but it is
1147 * possible for a page to have PageDirty set, but it is actually
1148 * clean (all its buffers are clean). This happens if the
1149 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001150 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001151 * try_to_release_page() will discover that cleanness and will
1152 * drop the buffers and mark the page clean - it can be freed.
1153 *
1154 * Rarely, pages can have buffers and no ->mapping. These are
1155 * the pages which were not successfully invalidated in
1156 * truncate_complete_page(). We try to drop those buffers here
1157 * and if that worked, and the page is no longer mapped into
1158 * process address space (page_count == 1) it can be freed.
1159 * Otherwise, leave the page on the LRU so it is swappable.
1160 */
David Howells266cf652009-04-03 16:42:36 +01001161 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001162 if (!try_to_release_page(page, sc->gfp_mask))
1163 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -07001164 if (!mapping && page_count(page) == 1) {
1165 unlock_page(page);
1166 if (put_page_testzero(page))
1167 goto free_it;
1168 else {
1169 /*
1170 * rare race with speculative reference.
1171 * the speculative reference will free
1172 * this page shortly, so we may
1173 * increment nr_reclaimed here (and
1174 * leave it off the LRU).
1175 */
1176 nr_reclaimed++;
1177 continue;
1178 }
1179 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001180 }
1181
Minchan Kim854e9ed2016-01-15 16:54:53 -08001182lazyfree:
Johannes Weinera5289102014-04-03 14:47:51 -07001183 if (!mapping || !__remove_mapping(mapping, page, true))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001184 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001185
Nick Piggina978d6f2008-10-18 20:26:58 -07001186 /*
1187 * At this point, we have no other references and there is
1188 * no way to pick any more up (removed from LRU, removed
1189 * from pagecache). Can use non-atomic bitops now (and
1190 * we obviously don't have to worry about waking up a process
1191 * waiting on the page lock, because there are no references.
1192 */
Kirill A. Shutemov48c935a2016-01-15 16:51:24 -08001193 __ClearPageLocked(page);
Nick Piggine2867812008-07-25 19:45:30 -07001194free_it:
Minchan Kim854e9ed2016-01-15 16:54:53 -08001195 if (ret == SWAP_LZFREE)
1196 count_vm_event(PGLAZYFREED);
1197
Andrew Morton05ff5132006-03-22 00:08:20 -08001198 nr_reclaimed++;
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001199
1200 /*
1201 * Is there need to periodically free_page_list? It would
1202 * appear not as the counts should be low
1203 */
1204 list_add(&page->lru, &free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001205 continue;
1206
Nick Pigginb291f002008-10-18 20:26:44 -07001207cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -08001208 if (PageSwapCache(page))
1209 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -07001210 unlock_page(page);
Jaewon Kimc54839a2015-09-08 15:02:21 -07001211 list_add(&page->lru, &ret_pages);
Nick Pigginb291f002008-10-18 20:26:44 -07001212 continue;
1213
Linus Torvalds1da177e2005-04-16 15:20:36 -07001214activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -07001215 /* Not a candidate for swapping, so reclaim swap space. */
Vladimir Davydov5ccc5ab2016-01-20 15:03:10 -08001216 if (PageSwapCache(page) && mem_cgroup_swap_full(page))
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08001217 try_to_free_swap(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001218 VM_BUG_ON_PAGE(PageActive(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001219 SetPageActive(page);
1220 pgactivate++;
1221keep_locked:
1222 unlock_page(page);
1223keep:
1224 list_add(&page->lru, &ret_pages);
Sasha Levin309381fea2014-01-23 15:52:54 -08001225 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001226 }
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001227
Johannes Weiner747db952014-08-08 14:19:24 -07001228 mem_cgroup_uncharge_list(&free_pages);
Mel Gorman72b252a2015-09-04 15:47:32 -07001229 try_to_unmap_flush();
Mel Gormanb745bc82014-06-04 16:10:22 -07001230 free_hot_cold_page_list(&free_pages, true);
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001231
Linus Torvalds1da177e2005-04-16 15:20:36 -07001232 list_splice(&ret_pages, page_list);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001233 count_vm_events(PGACTIVATE, pgactivate);
Johannes Weiner0a31bc92014-08-08 14:19:22 -07001234
Mel Gorman8e950282013-07-03 15:02:02 -07001235 *ret_nr_dirty += nr_dirty;
1236 *ret_nr_congested += nr_congested;
Mel Gormand43006d2013-07-03 15:01:50 -07001237 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
Mel Gorman92df3a72011-10-31 17:07:56 -07001238 *ret_nr_writeback += nr_writeback;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001239 *ret_nr_immediate += nr_immediate;
Andrew Morton05ff5132006-03-22 00:08:20 -08001240 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001241}
1242
Minchan Kim02c6de82012-10-08 16:31:55 -07001243unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1244 struct list_head *page_list)
1245{
1246 struct scan_control sc = {
1247 .gfp_mask = GFP_KERNEL,
1248 .priority = DEF_PRIORITY,
1249 .may_unmap = 1,
1250 };
Mel Gorman8e950282013-07-03 15:02:02 -07001251 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
Minchan Kim02c6de82012-10-08 16:31:55 -07001252 struct page *page, *next;
1253 LIST_HEAD(clean_pages);
1254
1255 list_for_each_entry_safe(page, next, page_list, lru) {
Rafael Aquini117aad12013-09-30 13:45:16 -07001256 if (page_is_file_cache(page) && !PageDirty(page) &&
1257 !isolated_balloon_page(page)) {
Minchan Kim02c6de82012-10-08 16:31:55 -07001258 ClearPageActive(page);
1259 list_move(&page->lru, &clean_pages);
1260 }
1261 }
1262
1263 ret = shrink_page_list(&clean_pages, zone, &sc,
Mel Gorman8e950282013-07-03 15:02:02 -07001264 TTU_UNMAP|TTU_IGNORE_ACCESS,
1265 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
Minchan Kim02c6de82012-10-08 16:31:55 -07001266 list_splice(&clean_pages, page_list);
Christoph Lameter83da7512014-04-18 15:07:10 -07001267 mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
Minchan Kim02c6de82012-10-08 16:31:55 -07001268 return ret;
1269}
1270
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001271/*
1272 * Attempt to remove the specified page from its LRU. Only take this page
1273 * if it is of the appropriate PageActive status. Pages which are being
1274 * freed elsewhere are also ignored.
1275 *
1276 * page: page to consider
1277 * mode: one of the LRU isolation modes defined above
1278 *
1279 * returns 0 on success, -ve errno on failure.
1280 */
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001281int __isolate_lru_page(struct page *page, isolate_mode_t mode)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001282{
1283 int ret = -EINVAL;
1284
1285 /* Only take pages on the LRU. */
1286 if (!PageLRU(page))
1287 return ret;
1288
Minchan Kime46a2872012-10-08 16:33:48 -07001289 /* Compaction should not handle unevictable pages but CMA can do so */
1290 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001291 return ret;
1292
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001293 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08001294
Mel Gormanc8244932012-01-12 17:19:38 -08001295 /*
1296 * To minimise LRU disruption, the caller can indicate that it only
1297 * wants to isolate pages it will be able to operate on without
1298 * blocking - clean pages for the most part.
1299 *
1300 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1301 * is used by reclaim when it is cannot write to backing storage
1302 *
1303 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1304 * that it is possible to migrate without blocking
1305 */
1306 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1307 /* All the caller can do on PageWriteback is block */
1308 if (PageWriteback(page))
1309 return ret;
1310
1311 if (PageDirty(page)) {
1312 struct address_space *mapping;
1313
1314 /* ISOLATE_CLEAN means only clean pages */
1315 if (mode & ISOLATE_CLEAN)
1316 return ret;
1317
1318 /*
1319 * Only pages without mappings or that have a
1320 * ->migratepage callback are possible to migrate
1321 * without blocking
1322 */
1323 mapping = page_mapping(page);
1324 if (mapping && !mapping->a_ops->migratepage)
1325 return ret;
1326 }
1327 }
Minchan Kim39deaf82011-10-31 17:06:51 -07001328
Minchan Kimf80c0672011-10-31 17:06:55 -07001329 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1330 return ret;
1331
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001332 if (likely(get_page_unless_zero(page))) {
1333 /*
1334 * Be careful not to clear PageLRU until after we're
1335 * sure the page is not being freed elsewhere -- the
1336 * page release code relies on it.
1337 */
1338 ClearPageLRU(page);
1339 ret = 0;
1340 }
1341
1342 return ret;
1343}
1344
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001345/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001346 * zone->lru_lock is heavily contended. Some of the functions that
1347 * shrink the lists perform better by taking out a batch of pages
1348 * and working on them outside the LRU lock.
1349 *
1350 * For pagecache intensive workloads, this function is the hottest
1351 * spot in the kernel (apart from copy_*_user functions).
1352 *
1353 * Appropriate locks must be held before calling this function.
1354 *
1355 * @nr_to_scan: The number of pages to look through on the list.
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001356 * @lruvec: The LRU vector to pull pages from.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001357 * @dst: The temp list to put pages on to.
Hugh Dickinsf6260122012-01-12 17:20:06 -08001358 * @nr_scanned: The number of pages that were scanned.
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001359 * @sc: The scan_control struct for this reclaim session
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001360 * @mode: One of the LRU isolation modes
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001361 * @lru: LRU list id for isolating
Linus Torvalds1da177e2005-04-16 15:20:36 -07001362 *
1363 * returns how many pages were moved onto *@dst.
1364 */
Andrew Morton69e05942006-03-22 00:08:19 -08001365static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001366 struct lruvec *lruvec, struct list_head *dst,
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001367 unsigned long *nr_scanned, struct scan_control *sc,
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001368 isolate_mode_t mode, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001369{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001370 struct list_head *src = &lruvec->lists[lru];
Andrew Morton69e05942006-03-22 00:08:19 -08001371 unsigned long nr_taken = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001372 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001373
Vladimir Davydov0b802f12015-09-08 15:02:18 -07001374 for (scan = 0; scan < nr_to_scan && nr_taken < nr_to_scan &&
1375 !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001376 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001377 int nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001378
Linus Torvalds1da177e2005-04-16 15:20:36 -07001379 page = lru_to_page(src);
1380 prefetchw_prev_lru_page(page, src, flags);
1381
Sasha Levin309381fea2014-01-23 15:52:54 -08001382 VM_BUG_ON_PAGE(!PageLRU(page), page);
Nick Piggin8d438f92006-03-22 00:07:59 -08001383
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001384 switch (__isolate_lru_page(page, mode)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001385 case 0:
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001386 nr_pages = hpage_nr_pages(page);
1387 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001388 list_move(&page->lru, dst);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001389 nr_taken += nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001390 break;
Nick Piggin46453a62006-03-22 00:07:58 -08001391
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001392 case -EBUSY:
1393 /* else it is being freed elsewhere */
1394 list_move(&page->lru, src);
1395 continue;
1396
1397 default:
1398 BUG();
1399 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001400 }
1401
Hugh Dickinsf6260122012-01-12 17:20:06 -08001402 *nr_scanned = scan;
Hugh Dickins75b00af2012-05-29 15:07:09 -07001403 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1404 nr_taken, mode, is_file_lru(lru));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001405 return nr_taken;
1406}
1407
Nick Piggin62695a82008-10-18 20:26:09 -07001408/**
1409 * isolate_lru_page - tries to isolate a page from its LRU list
1410 * @page: page to isolate from its LRU list
1411 *
1412 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1413 * vmstat statistic corresponding to whatever LRU list the page was on.
1414 *
1415 * Returns 0 if the page was removed from an LRU list.
1416 * Returns -EBUSY if the page was not on an LRU list.
1417 *
1418 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001419 * the active list, it will have PageActive set. If it was found on
1420 * the unevictable list, it will have the PageUnevictable bit set. That flag
1421 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001422 *
1423 * The vmstat statistic corresponding to the list on which the page was
1424 * found will be decremented.
1425 *
1426 * Restrictions:
1427 * (1) Must be called with an elevated refcount on the page. This is a
1428 * fundamentnal difference from isolate_lru_pages (which is called
1429 * without a stable reference).
1430 * (2) the lru_lock must not be held.
1431 * (3) interrupts must be enabled.
1432 */
1433int isolate_lru_page(struct page *page)
1434{
1435 int ret = -EBUSY;
1436
Sasha Levin309381fea2014-01-23 15:52:54 -08001437 VM_BUG_ON_PAGE(!page_count(page), page);
Kirill A. Shutemovcf2a82e2016-02-05 15:36:36 -08001438 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001439
Nick Piggin62695a82008-10-18 20:26:09 -07001440 if (PageLRU(page)) {
1441 struct zone *zone = page_zone(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001442 struct lruvec *lruvec;
Nick Piggin62695a82008-10-18 20:26:09 -07001443
1444 spin_lock_irq(&zone->lru_lock);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001445 lruvec = mem_cgroup_page_lruvec(page, zone);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001446 if (PageLRU(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001447 int lru = page_lru(page);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001448 get_page(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001449 ClearPageLRU(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001450 del_page_from_lru_list(page, lruvec, lru);
1451 ret = 0;
Nick Piggin62695a82008-10-18 20:26:09 -07001452 }
1453 spin_unlock_irq(&zone->lru_lock);
1454 }
1455 return ret;
1456}
1457
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001458/*
Fengguang Wud37dd5d2012-12-18 14:23:28 -08001459 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1460 * then get resheduled. When there are massive number of tasks doing page
1461 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1462 * the LRU list will go small and be scanned faster than necessary, leading to
1463 * unnecessary swapping, thrashing and OOM.
Rik van Riel35cd7812009-09-21 17:01:38 -07001464 */
1465static int too_many_isolated(struct zone *zone, int file,
1466 struct scan_control *sc)
1467{
1468 unsigned long inactive, isolated;
1469
1470 if (current_is_kswapd())
1471 return 0;
1472
Tejun Heo97c93412015-05-22 18:23:36 -04001473 if (!sane_reclaim(sc))
Rik van Riel35cd7812009-09-21 17:01:38 -07001474 return 0;
1475
1476 if (file) {
1477 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1478 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1479 } else {
1480 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1481 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1482 }
1483
Fengguang Wu3cf23842012-12-18 14:23:31 -08001484 /*
1485 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1486 * won't get blocked by normal direct-reclaimers, forming a circular
1487 * deadlock.
1488 */
Mel Gormand0164ad2015-11-06 16:28:21 -08001489 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
Fengguang Wu3cf23842012-12-18 14:23:31 -08001490 inactive >>= 3;
1491
Rik van Riel35cd7812009-09-21 17:01:38 -07001492 return isolated > inactive;
1493}
1494
Mel Gorman66635622010-08-09 17:19:30 -07001495static noinline_for_stack void
Hugh Dickins75b00af2012-05-29 15:07:09 -07001496putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
Mel Gorman66635622010-08-09 17:19:30 -07001497{
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001498 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1499 struct zone *zone = lruvec_zone(lruvec);
Hugh Dickins3f797682012-01-12 17:20:07 -08001500 LIST_HEAD(pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001501
Mel Gorman66635622010-08-09 17:19:30 -07001502 /*
1503 * Put back any unfreeable pages.
1504 */
Mel Gorman66635622010-08-09 17:19:30 -07001505 while (!list_empty(page_list)) {
Hugh Dickins3f797682012-01-12 17:20:07 -08001506 struct page *page = lru_to_page(page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001507 int lru;
Hugh Dickins3f797682012-01-12 17:20:07 -08001508
Sasha Levin309381fea2014-01-23 15:52:54 -08001509 VM_BUG_ON_PAGE(PageLRU(page), page);
Mel Gorman66635622010-08-09 17:19:30 -07001510 list_del(&page->lru);
Hugh Dickins39b5f292012-10-08 16:33:18 -07001511 if (unlikely(!page_evictable(page))) {
Mel Gorman66635622010-08-09 17:19:30 -07001512 spin_unlock_irq(&zone->lru_lock);
1513 putback_lru_page(page);
1514 spin_lock_irq(&zone->lru_lock);
1515 continue;
1516 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001517
1518 lruvec = mem_cgroup_page_lruvec(page, zone);
1519
Linus Torvalds7a608572011-01-17 14:42:19 -08001520 SetPageLRU(page);
Mel Gorman66635622010-08-09 17:19:30 -07001521 lru = page_lru(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001522 add_page_to_lru_list(page, lruvec, lru);
1523
Mel Gorman66635622010-08-09 17:19:30 -07001524 if (is_active_lru(lru)) {
1525 int file = is_file_lru(lru);
Rik van Riel9992af12011-01-13 15:47:13 -08001526 int numpages = hpage_nr_pages(page);
1527 reclaim_stat->recent_rotated[file] += numpages;
Mel Gorman66635622010-08-09 17:19:30 -07001528 }
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001529 if (put_page_testzero(page)) {
1530 __ClearPageLRU(page);
1531 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001532 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001533
1534 if (unlikely(PageCompound(page))) {
1535 spin_unlock_irq(&zone->lru_lock);
Johannes Weiner747db952014-08-08 14:19:24 -07001536 mem_cgroup_uncharge(page);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001537 (*get_compound_page_dtor(page))(page);
1538 spin_lock_irq(&zone->lru_lock);
1539 } else
1540 list_add(&page->lru, &pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001541 }
1542 }
Mel Gorman66635622010-08-09 17:19:30 -07001543
Hugh Dickins3f797682012-01-12 17:20:07 -08001544 /*
1545 * To save our caller's stack, now use input list for pages to free.
1546 */
1547 list_splice(&pages_to_free, page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001548}
1549
1550/*
NeilBrown399ba0b2014-06-04 16:07:42 -07001551 * If a kernel thread (such as nfsd for loop-back mounts) services
1552 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1553 * In that case we should only throttle if the backing device it is
1554 * writing to is congested. In other cases it is safe to throttle.
1555 */
1556static int current_may_throttle(void)
1557{
1558 return !(current->flags & PF_LESS_THROTTLE) ||
1559 current->backing_dev_info == NULL ||
1560 bdi_write_congested(current->backing_dev_info);
1561}
1562
1563/*
Andrew Morton1742f192006-03-22 00:08:21 -08001564 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1565 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001566 */
Mel Gorman66635622010-08-09 17:19:30 -07001567static noinline_for_stack unsigned long
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001568shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001569 struct scan_control *sc, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001570{
1571 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001572 unsigned long nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001573 unsigned long nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001574 unsigned long nr_taken;
Mel Gorman8e950282013-07-03 15:02:02 -07001575 unsigned long nr_dirty = 0;
1576 unsigned long nr_congested = 0;
Mel Gormane2be15f2013-07-03 15:01:57 -07001577 unsigned long nr_unqueued_dirty = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -07001578 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001579 unsigned long nr_immediate = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001580 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001581 int file = is_file_lru(lru);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001582 struct zone *zone = lruvec_zone(lruvec);
1583 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001584
Rik van Riel35cd7812009-09-21 17:01:38 -07001585 while (unlikely(too_many_isolated(zone, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001586 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001587
1588 /* We are about to die and free our memory. Return now. */
1589 if (fatal_signal_pending(current))
1590 return SWAP_CLUSTER_MAX;
1591 }
1592
Linus Torvalds1da177e2005-04-16 15:20:36 -07001593 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001594
1595 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001596 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001597 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001598 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001599
Linus Torvalds1da177e2005-04-16 15:20:36 -07001600 spin_lock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001601
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001602 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1603 &nr_scanned, sc, isolate_mode, lru);
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001604
1605 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1606 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1607
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001608 if (global_reclaim(sc)) {
Mel Gorman0d5d8232014-08-06 16:07:16 -07001609 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001610 if (current_is_kswapd())
Hugh Dickins75b00af2012-05-29 15:07:09 -07001611 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001612 else
Hugh Dickins75b00af2012-05-29 15:07:09 -07001613 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001614 }
Hillf Dantond563c052012-03-21 16:34:02 -07001615 spin_unlock_irq(&zone->lru_lock);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001616
Hillf Dantond563c052012-03-21 16:34:02 -07001617 if (nr_taken == 0)
Mel Gorman66635622010-08-09 17:19:30 -07001618 return 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001619
Minchan Kim02c6de82012-10-08 16:31:55 -07001620 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
Mel Gorman8e950282013-07-03 15:02:02 -07001621 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1622 &nr_writeback, &nr_immediate,
1623 false);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001624
Hugh Dickins3f797682012-01-12 17:20:07 -08001625 spin_lock_irq(&zone->lru_lock);
1626
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001627 reclaim_stat->recent_scanned[file] += nr_taken;
Hillf Dantond563c052012-03-21 16:34:02 -07001628
Ying Han904249a2012-04-25 16:01:48 -07001629 if (global_reclaim(sc)) {
1630 if (current_is_kswapd())
1631 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1632 nr_reclaimed);
1633 else
1634 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1635 nr_reclaimed);
1636 }
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001637
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001638 putback_inactive_pages(lruvec, &page_list);
Hugh Dickins3f797682012-01-12 17:20:07 -08001639
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001640 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Hugh Dickins3f797682012-01-12 17:20:07 -08001641
1642 spin_unlock_irq(&zone->lru_lock);
1643
Johannes Weiner747db952014-08-08 14:19:24 -07001644 mem_cgroup_uncharge_list(&page_list);
Mel Gormanb745bc82014-06-04 16:10:22 -07001645 free_hot_cold_page_list(&page_list, true);
Mel Gormane11da5b2010-10-26 14:21:40 -07001646
Mel Gorman92df3a72011-10-31 17:07:56 -07001647 /*
1648 * If reclaim is isolating dirty pages under writeback, it implies
1649 * that the long-lived page allocation rate is exceeding the page
1650 * laundering rate. Either the global limits are not being effective
1651 * at throttling processes due to the page distribution throughout
1652 * zones or there is heavy usage of a slow backing device. The
1653 * only option is to throttle from reclaim context which is not ideal
1654 * as there is no guarantee the dirtying process is throttled in the
1655 * same way balance_dirty_pages() manages.
1656 *
Mel Gorman8e950282013-07-03 15:02:02 -07001657 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1658 * of pages under pages flagged for immediate reclaim and stall if any
1659 * are encountered in the nr_immediate check below.
Mel Gorman92df3a72011-10-31 17:07:56 -07001660 */
Mel Gorman918fc712013-07-08 16:00:25 -07001661 if (nr_writeback && nr_writeback == nr_taken)
Johannes Weiner57054652014-10-09 15:28:17 -07001662 set_bit(ZONE_WRITEBACK, &zone->flags);
Mel Gorman92df3a72011-10-31 17:07:56 -07001663
Mel Gormand43006d2013-07-03 15:01:50 -07001664 /*
Tejun Heo97c93412015-05-22 18:23:36 -04001665 * Legacy memcg will stall in page writeback so avoid forcibly
1666 * stalling here.
Mel Gormand43006d2013-07-03 15:01:50 -07001667 */
Tejun Heo97c93412015-05-22 18:23:36 -04001668 if (sane_reclaim(sc)) {
Mel Gormanb1a6f212013-07-03 15:01:58 -07001669 /*
Mel Gorman8e950282013-07-03 15:02:02 -07001670 * Tag a zone as congested if all the dirty pages scanned were
1671 * backed by a congested BDI and wait_iff_congested will stall.
1672 */
1673 if (nr_dirty && nr_dirty == nr_congested)
Johannes Weiner57054652014-10-09 15:28:17 -07001674 set_bit(ZONE_CONGESTED, &zone->flags);
Mel Gorman8e950282013-07-03 15:02:02 -07001675
1676 /*
Mel Gormanb1a6f212013-07-03 15:01:58 -07001677 * If dirty pages are scanned that are not queued for IO, it
1678 * implies that flushers are not keeping up. In this case, flag
Johannes Weiner57054652014-10-09 15:28:17 -07001679 * the zone ZONE_DIRTY and kswapd will start writing pages from
1680 * reclaim context.
Mel Gormanb1a6f212013-07-03 15:01:58 -07001681 */
1682 if (nr_unqueued_dirty == nr_taken)
Johannes Weiner57054652014-10-09 15:28:17 -07001683 set_bit(ZONE_DIRTY, &zone->flags);
Mel Gormanb1a6f212013-07-03 15:01:58 -07001684
1685 /*
Linus Torvaldsb738d762014-06-08 14:17:00 -07001686 * If kswapd scans pages marked marked for immediate
1687 * reclaim and under writeback (nr_immediate), it implies
1688 * that pages are cycling through the LRU faster than
Mel Gormanb1a6f212013-07-03 15:01:58 -07001689 * they are written so also forcibly stall.
1690 */
Linus Torvaldsb738d762014-06-08 14:17:00 -07001691 if (nr_immediate && current_may_throttle())
Mel Gormanb1a6f212013-07-03 15:01:58 -07001692 congestion_wait(BLK_RW_ASYNC, HZ/10);
Mel Gormane2be15f2013-07-03 15:01:57 -07001693 }
Mel Gormand43006d2013-07-03 15:01:50 -07001694
Mel Gorman8e950282013-07-03 15:02:02 -07001695 /*
1696 * Stall direct reclaim for IO completions if underlying BDIs or zone
1697 * is congested. Allow kswapd to continue until it starts encountering
1698 * unqueued dirty pages or cycling through the LRU too quickly.
1699 */
NeilBrown399ba0b2014-06-04 16:07:42 -07001700 if (!sc->hibernation_mode && !current_is_kswapd() &&
1701 current_may_throttle())
Mel Gorman8e950282013-07-03 15:02:02 -07001702 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1703
yalin wangba5e9572016-01-14 15:18:48 -08001704 trace_mm_vmscan_lru_shrink_inactive(zone, nr_scanned, nr_reclaimed,
1705 sc->priority, file);
Andrew Morton05ff5132006-03-22 00:08:20 -08001706 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001707}
1708
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001709/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001710 * This moves pages from the active list to the inactive list.
1711 *
1712 * We move them the other way if the page is referenced by one or more
1713 * processes, from rmap.
1714 *
1715 * If the pages are mostly unmapped, the processing is fast and it is
1716 * appropriate to hold zone->lru_lock across the whole operation. But if
1717 * the pages are mapped, the processing is slow (page_referenced()) so we
1718 * should drop zone->lru_lock around each page. It's impossible to balance
1719 * this, so instead we remove the pages from the LRU while processing them.
1720 * It is safe to rely on PG_active against the non-LRU pages in here because
1721 * nobody will play with that bit on a non-LRU page.
1722 *
1723 * The downside is that we have to touch page->_count against each page.
1724 * But we had to alter page->flags anyway.
1725 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001726
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001727static void move_active_pages_to_lru(struct lruvec *lruvec,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001728 struct list_head *list,
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001729 struct list_head *pages_to_free,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001730 enum lru_list lru)
1731{
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001732 struct zone *zone = lruvec_zone(lruvec);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001733 unsigned long pgmoved = 0;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001734 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001735 int nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001736
Wu Fengguang3eb41402009-06-16 15:33:13 -07001737 while (!list_empty(list)) {
1738 page = lru_to_page(list);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001739 lruvec = mem_cgroup_page_lruvec(page, zone);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001740
Sasha Levin309381fea2014-01-23 15:52:54 -08001741 VM_BUG_ON_PAGE(PageLRU(page), page);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001742 SetPageLRU(page);
1743
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001744 nr_pages = hpage_nr_pages(page);
1745 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
Johannes Weiner925b7672012-01-12 17:18:15 -08001746 list_move(&page->lru, &lruvec->lists[lru]);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001747 pgmoved += nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001748
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001749 if (put_page_testzero(page)) {
1750 __ClearPageLRU(page);
1751 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001752 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001753
1754 if (unlikely(PageCompound(page))) {
1755 spin_unlock_irq(&zone->lru_lock);
Johannes Weiner747db952014-08-08 14:19:24 -07001756 mem_cgroup_uncharge(page);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001757 (*get_compound_page_dtor(page))(page);
1758 spin_lock_irq(&zone->lru_lock);
1759 } else
1760 list_add(&page->lru, pages_to_free);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001761 }
1762 }
1763 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1764 if (!is_active_lru(lru))
1765 __count_vm_events(PGDEACTIVATE, pgmoved);
1766}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001767
Hugh Dickinsf6260122012-01-12 17:20:06 -08001768static void shrink_active_list(unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001769 struct lruvec *lruvec,
Johannes Weinerf16015f2012-01-12 17:17:52 -08001770 struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001771 enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001772{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001773 unsigned long nr_taken;
Hugh Dickinsf6260122012-01-12 17:20:06 -08001774 unsigned long nr_scanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001775 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001776 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001777 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001778 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001779 struct page *page;
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001780 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001781 unsigned long nr_rotated = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001782 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001783 int file = is_file_lru(lru);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001784 struct zone *zone = lruvec_zone(lruvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001785
1786 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001787
1788 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001789 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001790 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001791 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001792
Linus Torvalds1da177e2005-04-16 15:20:36 -07001793 spin_lock_irq(&zone->lru_lock);
Johannes Weiner925b7672012-01-12 17:18:15 -08001794
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001795 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1796 &nr_scanned, sc, isolate_mode, lru);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001797 if (global_reclaim(sc))
Mel Gorman0d5d8232014-08-06 16:07:16 -07001798 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001799
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001800 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001801
Hugh Dickinsf6260122012-01-12 17:20:06 -08001802 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001803 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001804 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001805 spin_unlock_irq(&zone->lru_lock);
1806
Linus Torvalds1da177e2005-04-16 15:20:36 -07001807 while (!list_empty(&l_hold)) {
1808 cond_resched();
1809 page = lru_to_page(&l_hold);
1810 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001811
Hugh Dickins39b5f292012-10-08 16:33:18 -07001812 if (unlikely(!page_evictable(page))) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001813 putback_lru_page(page);
1814 continue;
1815 }
1816
Mel Gormancc715d92012-03-21 16:34:00 -07001817 if (unlikely(buffer_heads_over_limit)) {
1818 if (page_has_private(page) && trylock_page(page)) {
1819 if (page_has_private(page))
1820 try_to_release_page(page, 0);
1821 unlock_page(page);
1822 }
1823 }
1824
Johannes Weinerc3ac9a82012-05-29 15:06:25 -07001825 if (page_referenced(page, 0, sc->target_mem_cgroup,
1826 &vm_flags)) {
Rik van Riel9992af12011-01-13 15:47:13 -08001827 nr_rotated += hpage_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07001828 /*
1829 * Identify referenced, file-backed active pages and
1830 * give them one more trip around the active list. So
1831 * that executable code get better chances to stay in
1832 * memory under moderate memory pressure. Anon pages
1833 * are not likely to be evicted by use-once streaming
1834 * IO, plus JVM can create lots of anon VM_EXEC pages,
1835 * so we ignore them here.
1836 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001837 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001838 list_add(&page->lru, &l_active);
1839 continue;
1840 }
1841 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001842
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001843 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001844 list_add(&page->lru, &l_inactive);
1845 }
1846
Andrew Mortonb5557492009-01-06 14:40:13 -08001847 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001848 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001849 */
Johannes Weiner2a1dc502008-12-01 03:00:35 +01001850 spin_lock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001851 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001852 * Count referenced pages from currently used mappings as rotated,
1853 * even though only some of them are actually re-activated. This
1854 * helps balance scan pressure between file and anonymous pages in
Jerome Marchand7c0db9e2014-08-06 16:08:01 -07001855 * get_scan_count.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001856 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001857 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001858
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001859 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1860 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001861 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001862 spin_unlock_irq(&zone->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001863
Johannes Weiner747db952014-08-08 14:19:24 -07001864 mem_cgroup_uncharge_list(&l_hold);
Mel Gormanb745bc82014-06-04 16:10:22 -07001865 free_hot_cold_page_list(&l_hold, true);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001866}
1867
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001868#ifdef CONFIG_SWAP
Yaowei Bai42e2e452015-11-05 18:47:36 -08001869static bool inactive_anon_is_low_global(struct zone *zone)
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001870{
1871 unsigned long active, inactive;
1872
1873 active = zone_page_state(zone, NR_ACTIVE_ANON);
1874 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1875
Yaowei Bai29d06bb2015-11-05 18:47:17 -08001876 return inactive * zone->inactive_ratio < active;
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001877}
1878
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001879/**
1880 * inactive_anon_is_low - check if anonymous pages need to be deactivated
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001881 * @lruvec: LRU vector to check
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001882 *
1883 * Returns true if the zone does not have enough inactive anon pages,
1884 * meaning some active anon pages need to be deactivated.
1885 */
Yaowei Bai42e2e452015-11-05 18:47:36 -08001886static bool inactive_anon_is_low(struct lruvec *lruvec)
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001887{
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001888 /*
1889 * If we don't have swap space, anonymous page deactivation
1890 * is pointless.
1891 */
1892 if (!total_swap_pages)
Yaowei Bai42e2e452015-11-05 18:47:36 -08001893 return false;
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001894
Hugh Dickinsc3c787e2012-05-29 15:06:52 -07001895 if (!mem_cgroup_disabled())
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001896 return mem_cgroup_inactive_anon_is_low(lruvec);
Johannes Weinerf16015f2012-01-12 17:17:52 -08001897
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001898 return inactive_anon_is_low_global(lruvec_zone(lruvec));
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001899}
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001900#else
Yaowei Bai42e2e452015-11-05 18:47:36 -08001901static inline bool inactive_anon_is_low(struct lruvec *lruvec)
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001902{
Yaowei Bai42e2e452015-11-05 18:47:36 -08001903 return false;
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001904}
1905#endif
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001906
Rik van Riel56e49d22009-06-16 15:32:28 -07001907/**
1908 * inactive_file_is_low - check if file pages need to be deactivated
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001909 * @lruvec: LRU vector to check
Rik van Riel56e49d22009-06-16 15:32:28 -07001910 *
1911 * When the system is doing streaming IO, memory pressure here
1912 * ensures that active file pages get deactivated, until more
1913 * than half of the file pages are on the inactive list.
1914 *
1915 * Once we get to that situation, protect the system's working
1916 * set from being evicted by disabling active file page aging.
1917 *
1918 * This uses a different ratio than the anonymous pages, because
1919 * the page cache uses a use-once replacement algorithm.
1920 */
Yaowei Bai42e2e452015-11-05 18:47:36 -08001921static bool inactive_file_is_low(struct lruvec *lruvec)
Rik van Riel56e49d22009-06-16 15:32:28 -07001922{
Johannes Weinere3790142013-02-22 16:35:19 -08001923 unsigned long inactive;
1924 unsigned long active;
Rik van Riel56e49d22009-06-16 15:32:28 -07001925
Johannes Weiner23047a92016-03-15 14:57:16 -07001926 inactive = lruvec_lru_size(lruvec, LRU_INACTIVE_FILE);
1927 active = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE);
Johannes Weinere3790142013-02-22 16:35:19 -08001928
1929 return active > inactive;
Rik van Riel56e49d22009-06-16 15:32:28 -07001930}
1931
Yaowei Bai42e2e452015-11-05 18:47:36 -08001932static bool inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
Rik van Rielb39415b2009-12-14 17:59:48 -08001933{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001934 if (is_file_lru(lru))
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001935 return inactive_file_is_low(lruvec);
Rik van Rielb39415b2009-12-14 17:59:48 -08001936 else
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001937 return inactive_anon_is_low(lruvec);
Rik van Rielb39415b2009-12-14 17:59:48 -08001938}
1939
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001940static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001941 struct lruvec *lruvec, struct scan_control *sc)
Christoph Lameterb69408e2008-10-18 20:26:14 -07001942{
Rik van Rielb39415b2009-12-14 17:59:48 -08001943 if (is_active_lru(lru)) {
Hugh Dickins75b00af2012-05-29 15:07:09 -07001944 if (inactive_list_is_low(lruvec, lru))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001945 shrink_active_list(nr_to_scan, lruvec, sc, lru);
Rik van Riel556adec2008-10-18 20:26:34 -07001946 return 0;
1947 }
1948
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001949 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001950}
1951
Johannes Weiner9a265112013-02-22 16:32:17 -08001952enum scan_balance {
1953 SCAN_EQUAL,
1954 SCAN_FRACT,
1955 SCAN_ANON,
1956 SCAN_FILE,
1957};
1958
Linus Torvalds1da177e2005-04-16 15:20:36 -07001959/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001960 * Determine how aggressively the anon and file LRU lists should be
1961 * scanned. The relative value of each set of LRU lists is determined
1962 * by looking at the fraction of the pages scanned we did rotate back
1963 * onto the active list instead of evict.
1964 *
Wanpeng Libe7bd592012-06-14 20:41:02 +08001965 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1966 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001967 */
Vladimir Davydov33377672016-01-20 15:02:59 -08001968static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
Johannes Weiner6b4f7792014-12-12 16:56:13 -08001969 struct scan_control *sc, unsigned long *nr,
1970 unsigned long *lru_pages)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001971{
Vladimir Davydov33377672016-01-20 15:02:59 -08001972 int swappiness = mem_cgroup_swappiness(memcg);
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001973 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Johannes Weiner9a265112013-02-22 16:32:17 -08001974 u64 fraction[2];
1975 u64 denominator = 0; /* gcc */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001976 struct zone *zone = lruvec_zone(lruvec);
Johannes Weiner9a265112013-02-22 16:32:17 -08001977 unsigned long anon_prio, file_prio;
1978 enum scan_balance scan_balance;
Johannes Weiner0bf14572014-04-08 16:04:10 -07001979 unsigned long anon, file;
Johannes Weiner9a265112013-02-22 16:32:17 -08001980 bool force_scan = false;
1981 unsigned long ap, fp;
1982 enum lru_list lru;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07001983 bool some_scanned;
1984 int pass;
KAMEZAWA Hiroyuki246e87a2011-05-26 16:25:34 -07001985
Johannes Weinerf11c0ca2011-10-31 17:07:27 -07001986 /*
1987 * If the zone or memcg is small, nr[l] can be 0. This
1988 * results in no scanning on this priority and a potential
1989 * priority drop. Global direct reclaim can go to the next
1990 * zone and tends to have no problems. Global kswapd is for
1991 * zone balancing and it needs to scan a minimum amount. When
1992 * reclaiming for a memcg, a priority drop can cause high
1993 * latencies, so it's better to scan a minimum amount there as
1994 * well.
1995 */
Vladimir Davydov90cbc252015-02-11 15:25:55 -08001996 if (current_is_kswapd()) {
1997 if (!zone_reclaimable(zone))
1998 force_scan = true;
Vladimir Davydoveb01aaa2016-01-20 15:03:02 -08001999 if (!mem_cgroup_online(memcg))
Vladimir Davydov90cbc252015-02-11 15:25:55 -08002000 force_scan = true;
2001 }
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002002 if (!global_reclaim(sc))
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07002003 force_scan = true;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002004
2005 /* If we have no swap space, do not bother scanning anon pages. */
Vladimir Davydovd8b38432016-01-20 15:03:07 -08002006 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002007 scan_balance = SCAN_FILE;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002008 goto out;
2009 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002010
Johannes Weiner10316b32013-02-22 16:32:14 -08002011 /*
2012 * Global reclaim will swap to prevent OOM even with no
2013 * swappiness, but memcg users want to use this knob to
2014 * disable swapping for individual groups completely when
2015 * using the memory controller's swap limit feature would be
2016 * too expensive.
2017 */
Johannes Weiner02695172014-08-06 16:06:17 -07002018 if (!global_reclaim(sc) && !swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002019 scan_balance = SCAN_FILE;
Johannes Weiner10316b32013-02-22 16:32:14 -08002020 goto out;
2021 }
2022
2023 /*
2024 * Do not apply any pressure balancing cleverness when the
2025 * system is close to OOM, scan both anon and file equally
2026 * (unless the swappiness setting disagrees with swapping).
2027 */
Johannes Weiner02695172014-08-06 16:06:17 -07002028 if (!sc->priority && swappiness) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002029 scan_balance = SCAN_EQUAL;
Johannes Weiner10316b32013-02-22 16:32:14 -08002030 goto out;
2031 }
2032
Johannes Weiner11d16c22013-02-22 16:32:15 -08002033 /*
Johannes Weiner62376252014-05-06 12:50:07 -07002034 * Prevent the reclaimer from falling into the cache trap: as
2035 * cache pages start out inactive, every cache fault will tip
2036 * the scan balance towards the file LRU. And as the file LRU
2037 * shrinks, so does the window for rotation from references.
2038 * This means we have a runaway feedback loop where a tiny
2039 * thrashing file LRU becomes infinitely more attractive than
2040 * anon pages. Try to detect this based on file LRU size.
2041 */
2042 if (global_reclaim(sc)) {
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002043 unsigned long zonefile;
2044 unsigned long zonefree;
Johannes Weiner62376252014-05-06 12:50:07 -07002045
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002046 zonefree = zone_page_state(zone, NR_FREE_PAGES);
2047 zonefile = zone_page_state(zone, NR_ACTIVE_FILE) +
2048 zone_page_state(zone, NR_INACTIVE_FILE);
2049
2050 if (unlikely(zonefile + zonefree <= high_wmark_pages(zone))) {
Johannes Weiner62376252014-05-06 12:50:07 -07002051 scan_balance = SCAN_ANON;
2052 goto out;
2053 }
2054 }
2055
2056 /*
Vladimir Davydov316bda02016-01-14 15:19:38 -08002057 * If there is enough inactive page cache, i.e. if the size of the
2058 * inactive list is greater than that of the active list *and* the
2059 * inactive list actually has some pages to scan on this priority, we
2060 * do not reclaim anything from the anonymous working set right now.
2061 * Without the second condition we could end up never scanning an
2062 * lruvec even if it has plenty of old anonymous pages unless the
2063 * system is under heavy pressure.
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002064 */
Vladimir Davydov316bda02016-01-14 15:19:38 -08002065 if (!inactive_file_is_low(lruvec) &&
Johannes Weiner23047a92016-03-15 14:57:16 -07002066 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE) >> sc->priority) {
Johannes Weiner9a265112013-02-22 16:32:17 -08002067 scan_balance = SCAN_FILE;
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002068 goto out;
2069 }
2070
Johannes Weiner9a265112013-02-22 16:32:17 -08002071 scan_balance = SCAN_FRACT;
2072
Johannes Weiner7c5bd702013-02-22 16:32:10 -08002073 /*
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07002074 * With swappiness at 100, anonymous and file have the same priority.
2075 * This scanning priority is essentially the inverse of IO cost.
2076 */
Johannes Weiner02695172014-08-06 16:06:17 -07002077 anon_prio = swappiness;
Hugh Dickins75b00af2012-05-29 15:07:09 -07002078 file_prio = 200 - anon_prio;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07002079
2080 /*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002081 * OK, so we have swap space and a fair amount of page cache
2082 * pages. We use the recently rotated / recently scanned
2083 * ratios to determine how valuable each cache is.
2084 *
2085 * Because workloads change over time (and to avoid overflow)
2086 * we keep these statistics as a floating average, which ends
2087 * up weighing recent references more than old ones.
2088 *
2089 * anon in [0], file in [1]
2090 */
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002091
Johannes Weiner23047a92016-03-15 14:57:16 -07002092 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON) +
2093 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON);
2094 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE) +
2095 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE);
Jerome Marchand2ab051e2014-08-06 16:08:03 -07002096
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07002097 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002098 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002099 reclaim_stat->recent_scanned[0] /= 2;
2100 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002101 }
2102
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002103 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002104 reclaim_stat->recent_scanned[1] /= 2;
2105 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002106 }
2107
2108 /*
Rik van Riel00d80892008-11-19 15:36:44 -08002109 * The amount of pressure on anon vs file pages is inversely
2110 * proportional to the fraction of recently scanned pages on
2111 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002112 */
Satoru Moriyafe350042012-05-29 15:06:47 -07002113 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002114 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002115
Satoru Moriyafe350042012-05-29 15:06:47 -07002116 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08002117 fp /= reclaim_stat->recent_rotated[1] + 1;
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07002118 spin_unlock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002119
Shaohua Li76a33fc2010-05-24 14:32:36 -07002120 fraction[0] = ap;
2121 fraction[1] = fp;
2122 denominator = ap + fp + 1;
2123out:
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002124 some_scanned = false;
2125 /* Only use force_scan on second pass. */
2126 for (pass = 0; !some_scanned && pass < 2; pass++) {
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002127 *lru_pages = 0;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002128 for_each_evictable_lru(lru) {
2129 int file = is_file_lru(lru);
2130 unsigned long size;
2131 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002132
Johannes Weiner23047a92016-03-15 14:57:16 -07002133 size = lruvec_lru_size(lruvec, lru);
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002134 scan = size >> sc->priority;
Johannes Weiner9a265112013-02-22 16:32:17 -08002135
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002136 if (!scan && pass && force_scan)
2137 scan = min(size, SWAP_CLUSTER_MAX);
Johannes Weiner9a265112013-02-22 16:32:17 -08002138
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002139 switch (scan_balance) {
2140 case SCAN_EQUAL:
2141 /* Scan lists relative to size */
2142 break;
2143 case SCAN_FRACT:
2144 /*
2145 * Scan types proportional to swappiness and
2146 * their relative recent reclaim efficiency.
2147 */
2148 scan = div64_u64(scan * fraction[file],
2149 denominator);
2150 break;
2151 case SCAN_FILE:
2152 case SCAN_ANON:
2153 /* Scan one type exclusively */
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002154 if ((scan_balance == SCAN_FILE) != file) {
2155 size = 0;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002156 scan = 0;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002157 }
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002158 break;
2159 default:
2160 /* Look ma, no brain */
2161 BUG();
2162 }
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002163
2164 *lru_pages += size;
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002165 nr[lru] = scan;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002166
Johannes Weiner9a265112013-02-22 16:32:17 -08002167 /*
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002168 * Skip the second pass and don't force_scan,
2169 * if we found something to scan.
Johannes Weiner9a265112013-02-22 16:32:17 -08002170 */
Suleiman Souhlal6f04f482014-06-04 16:06:44 -07002171 some_scanned |= !!scan;
Johannes Weiner9a265112013-02-22 16:32:17 -08002172 }
Shaohua Li76a33fc2010-05-24 14:32:36 -07002173 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07002174}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002175
Mel Gorman72b252a2015-09-04 15:47:32 -07002176#ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
2177static void init_tlb_ubc(void)
2178{
2179 /*
2180 * This deliberately does not clear the cpumask as it's expensive
2181 * and unnecessary. If there happens to be data in there then the
2182 * first SWAP_CLUSTER_MAX pages will send an unnecessary IPI and
2183 * then will be cleared.
2184 */
2185 current->tlb_ubc.flush_required = false;
2186}
2187#else
2188static inline void init_tlb_ubc(void)
2189{
2190}
2191#endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
2192
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002193/*
2194 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2195 */
Vladimir Davydov33377672016-01-20 15:02:59 -08002196static void shrink_zone_memcg(struct zone *zone, struct mem_cgroup *memcg,
2197 struct scan_control *sc, unsigned long *lru_pages)
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002198{
Vladimir Davydov33377672016-01-20 15:02:59 -08002199 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002200 unsigned long nr[NR_LRU_LISTS];
Mel Gormane82e0562013-07-03 15:01:44 -07002201 unsigned long targets[NR_LRU_LISTS];
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002202 unsigned long nr_to_scan;
2203 enum lru_list lru;
2204 unsigned long nr_reclaimed = 0;
2205 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2206 struct blk_plug plug;
Mel Gorman1a501902014-06-04 16:10:49 -07002207 bool scan_adjusted;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002208
Vladimir Davydov33377672016-01-20 15:02:59 -08002209 get_scan_count(lruvec, memcg, sc, nr, lru_pages);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002210
Mel Gormane82e0562013-07-03 15:01:44 -07002211 /* Record the original scan target for proportional adjustments later */
2212 memcpy(targets, nr, sizeof(nr));
2213
Mel Gorman1a501902014-06-04 16:10:49 -07002214 /*
2215 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2216 * event that can occur when there is little memory pressure e.g.
2217 * multiple streaming readers/writers. Hence, we do not abort scanning
2218 * when the requested number of pages are reclaimed when scanning at
2219 * DEF_PRIORITY on the assumption that the fact we are direct
2220 * reclaiming implies that kswapd is not keeping up and it is best to
2221 * do a batch of work at once. For memcg reclaim one check is made to
2222 * abort proportional reclaim if either the file or anon lru has already
2223 * dropped to zero at the first pass.
2224 */
2225 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2226 sc->priority == DEF_PRIORITY);
2227
Mel Gorman72b252a2015-09-04 15:47:32 -07002228 init_tlb_ubc();
2229
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002230 blk_start_plug(&plug);
2231 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2232 nr[LRU_INACTIVE_FILE]) {
Mel Gormane82e0562013-07-03 15:01:44 -07002233 unsigned long nr_anon, nr_file, percentage;
2234 unsigned long nr_scanned;
2235
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002236 for_each_evictable_lru(lru) {
2237 if (nr[lru]) {
2238 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2239 nr[lru] -= nr_to_scan;
2240
2241 nr_reclaimed += shrink_list(lru, nr_to_scan,
2242 lruvec, sc);
2243 }
2244 }
Mel Gormane82e0562013-07-03 15:01:44 -07002245
2246 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2247 continue;
2248
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002249 /*
Mel Gormane82e0562013-07-03 15:01:44 -07002250 * For kswapd and memcg, reclaim at least the number of pages
Mel Gorman1a501902014-06-04 16:10:49 -07002251 * requested. Ensure that the anon and file LRUs are scanned
Mel Gormane82e0562013-07-03 15:01:44 -07002252 * proportionally what was requested by get_scan_count(). We
2253 * stop reclaiming one LRU and reduce the amount scanning
2254 * proportional to the original scan target.
2255 */
2256 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2257 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2258
Mel Gorman1a501902014-06-04 16:10:49 -07002259 /*
2260 * It's just vindictive to attack the larger once the smaller
2261 * has gone to zero. And given the way we stop scanning the
2262 * smaller below, this makes sure that we only make one nudge
2263 * towards proportionality once we've got nr_to_reclaim.
2264 */
2265 if (!nr_file || !nr_anon)
2266 break;
2267
Mel Gormane82e0562013-07-03 15:01:44 -07002268 if (nr_file > nr_anon) {
2269 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2270 targets[LRU_ACTIVE_ANON] + 1;
2271 lru = LRU_BASE;
2272 percentage = nr_anon * 100 / scan_target;
2273 } else {
2274 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2275 targets[LRU_ACTIVE_FILE] + 1;
2276 lru = LRU_FILE;
2277 percentage = nr_file * 100 / scan_target;
2278 }
2279
2280 /* Stop scanning the smaller of the LRU */
2281 nr[lru] = 0;
2282 nr[lru + LRU_ACTIVE] = 0;
2283
2284 /*
2285 * Recalculate the other LRU scan count based on its original
2286 * scan target and the percentage scanning already complete
2287 */
2288 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2289 nr_scanned = targets[lru] - nr[lru];
2290 nr[lru] = targets[lru] * (100 - percentage) / 100;
2291 nr[lru] -= min(nr[lru], nr_scanned);
2292
2293 lru += LRU_ACTIVE;
2294 nr_scanned = targets[lru] - nr[lru];
2295 nr[lru] = targets[lru] * (100 - percentage) / 100;
2296 nr[lru] -= min(nr[lru], nr_scanned);
2297
2298 scan_adjusted = true;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002299 }
2300 blk_finish_plug(&plug);
2301 sc->nr_reclaimed += nr_reclaimed;
2302
2303 /*
2304 * Even if we did not try to evict anon pages at all, we want to
2305 * rebalance the anon lru active/inactive ratio.
2306 */
2307 if (inactive_anon_is_low(lruvec))
2308 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2309 sc, LRU_ACTIVE_ANON);
2310
2311 throttle_vm_writeout(sc->gfp_mask);
2312}
2313
Mel Gorman23b9da52012-05-29 15:06:20 -07002314/* Use reclaim/compaction for costly allocs or under memory pressure */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002315static bool in_reclaim_compaction(struct scan_control *sc)
Mel Gorman23b9da52012-05-29 15:06:20 -07002316{
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002317 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
Mel Gorman23b9da52012-05-29 15:06:20 -07002318 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002319 sc->priority < DEF_PRIORITY - 2))
Mel Gorman23b9da52012-05-29 15:06:20 -07002320 return true;
2321
2322 return false;
2323}
2324
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002325/*
Mel Gorman23b9da52012-05-29 15:06:20 -07002326 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2327 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2328 * true if more pages should be reclaimed such that when the page allocator
2329 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2330 * It will give up earlier than that if there is difficulty reclaiming pages.
Mel Gorman3e7d3442011-01-13 15:45:56 -08002331 */
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002332static inline bool should_continue_reclaim(struct zone *zone,
Mel Gorman3e7d3442011-01-13 15:45:56 -08002333 unsigned long nr_reclaimed,
2334 unsigned long nr_scanned,
2335 struct scan_control *sc)
2336{
2337 unsigned long pages_for_compaction;
2338 unsigned long inactive_lru_pages;
2339
2340 /* If not in reclaim/compaction mode, stop */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002341 if (!in_reclaim_compaction(sc))
Mel Gorman3e7d3442011-01-13 15:45:56 -08002342 return false;
2343
Mel Gorman28765922011-02-25 14:44:20 -08002344 /* Consider stopping depending on scan and reclaim activity */
2345 if (sc->gfp_mask & __GFP_REPEAT) {
2346 /*
2347 * For __GFP_REPEAT allocations, stop reclaiming if the
2348 * full LRU list has been scanned and we are still failing
2349 * to reclaim pages. This full LRU scan is potentially
2350 * expensive but a __GFP_REPEAT caller really wants to succeed
2351 */
2352 if (!nr_reclaimed && !nr_scanned)
2353 return false;
2354 } else {
2355 /*
2356 * For non-__GFP_REPEAT allocations which can presumably
2357 * fail without consequence, stop if we failed to reclaim
2358 * any pages from the last SWAP_CLUSTER_MAX number of
2359 * pages that were scanned. This will return to the
2360 * caller faster at the risk reclaim/compaction and
2361 * the resulting allocation attempt fails
2362 */
2363 if (!nr_reclaimed)
2364 return false;
2365 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002366
2367 /*
2368 * If we have not reclaimed enough pages for compaction and the
2369 * inactive lists are large enough, continue reclaiming
2370 */
2371 pages_for_compaction = (2UL << sc->order);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002372 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
Shaohua Liec8acf22013-02-22 16:34:38 -08002373 if (get_nr_swap_pages() > 0)
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002374 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
Mel Gorman3e7d3442011-01-13 15:45:56 -08002375 if (sc->nr_reclaimed < pages_for_compaction &&
2376 inactive_lru_pages > pages_for_compaction)
2377 return true;
2378
2379 /* If compaction would go ahead or the allocation would succeed, stop */
Vlastimil Babkaebff3982014-12-10 15:43:22 -08002380 switch (compaction_suitable(zone, sc->order, 0, 0)) {
Mel Gorman3e7d3442011-01-13 15:45:56 -08002381 case COMPACT_PARTIAL:
2382 case COMPACT_CONTINUE:
2383 return false;
2384 default:
2385 return true;
2386 }
2387}
2388
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002389static bool shrink_zone(struct zone *zone, struct scan_control *sc,
2390 bool is_classzone)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002391{
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002392 struct reclaim_state *reclaim_state = current->reclaim_state;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002393 unsigned long nr_reclaimed, nr_scanned;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002394 bool reclaimable = false;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002395
Johannes Weiner56600482012-01-12 17:17:59 -08002396 do {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002397 struct mem_cgroup *root = sc->target_mem_cgroup;
2398 struct mem_cgroup_reclaim_cookie reclaim = {
2399 .zone = zone,
2400 .priority = sc->priority,
2401 };
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002402 unsigned long zone_lru_pages = 0;
Andrew Morton694fbc02013-09-24 15:27:37 -07002403 struct mem_cgroup *memcg;
Johannes Weiner56600482012-01-12 17:17:59 -08002404
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002405 nr_reclaimed = sc->nr_reclaimed;
2406 nr_scanned = sc->nr_scanned;
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002407
Andrew Morton694fbc02013-09-24 15:27:37 -07002408 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2409 do {
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002410 unsigned long lru_pages;
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002411 unsigned long reclaimed;
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002412 unsigned long scanned;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002413
Johannes Weiner241994ed2015-02-11 15:26:06 -08002414 if (mem_cgroup_low(root, memcg)) {
2415 if (!sc->may_thrash)
2416 continue;
2417 mem_cgroup_events(memcg, MEMCG_LOW, 1);
2418 }
2419
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002420 reclaimed = sc->nr_reclaimed;
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002421 scanned = sc->nr_scanned;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002422
Vladimir Davydov33377672016-01-20 15:02:59 -08002423 shrink_zone_memcg(zone, memcg, sc, &lru_pages);
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002424 zone_lru_pages += lru_pages;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002425
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002426 if (memcg && is_classzone)
2427 shrink_slab(sc->gfp_mask, zone_to_nid(zone),
2428 memcg, sc->nr_scanned - scanned,
2429 lru_pages);
2430
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002431 /* Record the group's reclaim efficiency */
2432 vmpressure(sc->gfp_mask, memcg, false,
2433 sc->nr_scanned - scanned,
2434 sc->nr_reclaimed - reclaimed);
2435
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002436 /*
Michal Hockoa394cb82013-02-22 16:32:30 -08002437 * Direct reclaim and kswapd have to scan all memory
2438 * cgroups to fulfill the overall scan target for the
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002439 * zone.
Michal Hockoa394cb82013-02-22 16:32:30 -08002440 *
2441 * Limit reclaim, on the other hand, only cares about
2442 * nr_to_reclaim pages to be reclaimed and it will
2443 * retry with decreasing priority if one round over the
2444 * whole hierarchy is not sufficient.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002445 */
Michal Hockoa394cb82013-02-22 16:32:30 -08002446 if (!global_reclaim(sc) &&
2447 sc->nr_reclaimed >= sc->nr_to_reclaim) {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002448 mem_cgroup_iter_break(root, memcg);
2449 break;
2450 }
Johannes Weiner241994ed2015-02-11 15:26:06 -08002451 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002452
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002453 /*
2454 * Shrink the slab caches in the same proportion that
2455 * the eligible LRU pages were scanned.
2456 */
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002457 if (global_reclaim(sc) && is_classzone)
2458 shrink_slab(sc->gfp_mask, zone_to_nid(zone), NULL,
2459 sc->nr_scanned - nr_scanned,
2460 zone_lru_pages);
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002461
Vladimir Davydovcb731d62015-02-12 14:58:54 -08002462 if (reclaim_state) {
2463 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2464 reclaim_state->reclaimed_slab = 0;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002465 }
2466
Johannes Weiner8e8ae642016-01-14 15:21:32 -08002467 /* Record the subtree's reclaim efficiency */
2468 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002469 sc->nr_scanned - nr_scanned,
2470 sc->nr_reclaimed - nr_reclaimed);
2471
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002472 if (sc->nr_reclaimed - nr_reclaimed)
2473 reclaimable = true;
2474
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002475 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2476 sc->nr_scanned - nr_scanned, sc));
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002477
2478 return reclaimable;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002479}
2480
Vlastimil Babka53853e22014-10-09 15:27:02 -07002481/*
2482 * Returns true if compaction should go ahead for a high-order request, or
2483 * the high-order allocation would succeed without compaction.
2484 */
Johannes Weiner0b064962014-08-06 16:06:12 -07002485static inline bool compaction_ready(struct zone *zone, int order)
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002486{
2487 unsigned long balance_gap, watermark;
2488 bool watermark_ok;
2489
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002490 /*
2491 * Compaction takes time to run and there are potentially other
2492 * callers using the pages just freed. Continue reclaiming until
2493 * there is a buffer of free pages available to give compaction
2494 * a reasonable chance of completing and allocating the page
2495 */
Jianyu Zhan4be89a32014-06-04 16:10:38 -07002496 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
2497 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
Johannes Weiner0b064962014-08-06 16:06:12 -07002498 watermark = high_wmark_pages(zone) + balance_gap + (2UL << order);
Mel Gormane2b19192015-11-06 16:28:09 -08002499 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0);
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002500
2501 /*
2502 * If compaction is deferred, reclaim up to a point where
2503 * compaction will have a chance of success when re-enabled
2504 */
Johannes Weiner0b064962014-08-06 16:06:12 -07002505 if (compaction_deferred(zone, order))
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002506 return watermark_ok;
2507
Vlastimil Babka53853e22014-10-09 15:27:02 -07002508 /*
2509 * If compaction is not ready to start and allocation is not likely
2510 * to succeed without it, then keep reclaiming.
2511 */
Vlastimil Babkaebff3982014-12-10 15:43:22 -08002512 if (compaction_suitable(zone, order, 0, 0) == COMPACT_SKIPPED)
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002513 return false;
2514
2515 return watermark_ok;
2516}
2517
Linus Torvalds1da177e2005-04-16 15:20:36 -07002518/*
2519 * This is the direct reclaim path, for page-allocating processes. We only
2520 * try to reclaim pages from zones which will satisfy the caller's allocation
2521 * request.
2522 *
Mel Gorman41858962009-06-16 15:32:12 -07002523 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2524 * Because:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002525 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2526 * allocation or
Mel Gorman41858962009-06-16 15:32:12 -07002527 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2528 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2529 * zone defense algorithm.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002530 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002531 * If a zone is deemed to be full of pinned pages then just give it a light
2532 * scan then give up on it.
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002533 *
2534 * Returns true if a zone was reclaimable.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002535 */
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002536static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002537{
Mel Gormandd1a2392008-04-28 02:12:17 -07002538 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002539 struct zone *zone;
Andrew Morton0608f432013-09-24 15:27:41 -07002540 unsigned long nr_soft_reclaimed;
2541 unsigned long nr_soft_scanned;
Weijie Yang619d0d762014-04-07 15:36:59 -07002542 gfp_t orig_mask;
Weijie Yang9bbc04e2014-04-07 15:37:00 -07002543 enum zone_type requested_highidx = gfp_zone(sc->gfp_mask);
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002544 bool reclaimable = false;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002545
Mel Gormancc715d92012-03-21 16:34:00 -07002546 /*
2547 * If the number of buffer_heads in the machine exceeds the maximum
2548 * allowed level, force direct reclaim to scan the highmem zone as
2549 * highmem pages could be pinning lowmem pages storing buffer_heads
2550 */
Weijie Yang619d0d762014-04-07 15:36:59 -07002551 orig_mask = sc->gfp_mask;
Mel Gormancc715d92012-03-21 16:34:00 -07002552 if (buffer_heads_over_limit)
2553 sc->gfp_mask |= __GFP_HIGHMEM;
2554
Mel Gormand4debc62010-08-09 17:19:29 -07002555 for_each_zone_zonelist_nodemask(zone, z, zonelist,
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002556 requested_highidx, sc->nodemask) {
2557 enum zone_type classzone_idx;
2558
Con Kolivasf3fe6512006-01-06 00:11:15 -08002559 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002560 continue;
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002561
2562 classzone_idx = requested_highidx;
2563 while (!populated_zone(zone->zone_pgdat->node_zones +
2564 classzone_idx))
2565 classzone_idx--;
2566
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002567 /*
2568 * Take care memory controller reclaiming has small influence
2569 * to global LRU.
2570 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002571 if (global_reclaim(sc)) {
Vladimir Davydov344736f2014-10-20 15:50:30 +04002572 if (!cpuset_zone_allowed(zone,
2573 GFP_KERNEL | __GFP_HARDWALL))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002574 continue;
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002575
Lisa Du6e543d52013-09-11 14:22:36 -07002576 if (sc->priority != DEF_PRIORITY &&
2577 !zone_reclaimable(zone))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002578 continue; /* Let kswapd poll it */
Johannes Weiner0b064962014-08-06 16:06:12 -07002579
2580 /*
2581 * If we already have plenty of memory free for
2582 * compaction in this zone, don't free any more.
2583 * Even though compaction is invoked for any
2584 * non-zero order, only frequent costly order
2585 * reclamation is disruptive enough to become a
2586 * noticeable problem, like transparent huge
2587 * page allocations.
2588 */
2589 if (IS_ENABLED(CONFIG_COMPACTION) &&
2590 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2591 zonelist_zone_idx(z) <= requested_highidx &&
2592 compaction_ready(zone, sc->order)) {
2593 sc->compaction_ready = true;
2594 continue;
Rik van Riele0887c12011-10-31 17:09:31 -07002595 }
Johannes Weiner0b064962014-08-06 16:06:12 -07002596
Andrew Morton0608f432013-09-24 15:27:41 -07002597 /*
2598 * This steals pages from memory cgroups over softlimit
2599 * and returns the number of reclaimed pages and
2600 * scanned pages. This works for global memory pressure
2601 * and balancing, not for a memcg's limit.
2602 */
2603 nr_soft_scanned = 0;
2604 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2605 sc->order, sc->gfp_mask,
2606 &nr_soft_scanned);
2607 sc->nr_reclaimed += nr_soft_reclaimed;
2608 sc->nr_scanned += nr_soft_scanned;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002609 if (nr_soft_reclaimed)
2610 reclaimable = true;
KAMEZAWA Hiroyukiac34a1a2011-06-27 16:18:12 -07002611 /* need some check for avoid more shrink_zone() */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002612 }
Nick Piggin408d8542006-09-25 23:31:27 -07002613
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002614 if (shrink_zone(zone, sc, zone_idx(zone) == classzone_idx))
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002615 reclaimable = true;
2616
2617 if (global_reclaim(sc) &&
2618 !reclaimable && zone_reclaimable(zone))
2619 reclaimable = true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002620 }
Mel Gormane0c23272011-10-31 17:09:33 -07002621
Vladimir Davydov65ec02c2014-04-03 14:47:20 -07002622 /*
Weijie Yang619d0d762014-04-07 15:36:59 -07002623 * Restore to original mask to avoid the impact on the caller if we
2624 * promoted it to __GFP_HIGHMEM.
2625 */
2626 sc->gfp_mask = orig_mask;
Minchan Kimd1908362010-09-22 13:05:01 -07002627
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002628 return reclaimable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002629}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002630
Linus Torvalds1da177e2005-04-16 15:20:36 -07002631/*
2632 * This is the main entry point to direct page reclaim.
2633 *
2634 * If a full scan of the inactive list fails to free enough memory then we
2635 * are "out of memory" and something needs to be killed.
2636 *
2637 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2638 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02002639 * caller can't do much about. We kick the writeback threads and take explicit
2640 * naps in the hope that some of these pages can be written. But if the
2641 * allocating task holds filesystem locks which prevent writeout this might not
2642 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07002643 *
2644 * returns: 0, if no pages reclaimed
2645 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07002646 */
Mel Gormandac1d272008-04-28 02:12:12 -07002647static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002648 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002649{
Johannes Weiner241994ed2015-02-11 15:26:06 -08002650 int initial_priority = sc->priority;
Andrew Morton69e05942006-03-22 00:08:19 -08002651 unsigned long total_scanned = 0;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002652 unsigned long writeback_threshold;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002653 bool zones_reclaimable;
Johannes Weiner241994ed2015-02-11 15:26:06 -08002654retry:
Keika Kobayashi873b4772008-07-25 01:48:52 -07002655 delayacct_freepages_start();
2656
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002657 if (global_reclaim(sc))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002658 count_vm_event(ALLOCSTALL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002659
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002660 do {
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002661 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2662 sc->priority);
Balbir Singh66e17072008-02-07 00:13:56 -08002663 sc->nr_scanned = 0;
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002664 zones_reclaimable = shrink_zones(zonelist, sc);
Mel Gormane0c23272011-10-31 17:09:33 -07002665
Balbir Singh66e17072008-02-07 00:13:56 -08002666 total_scanned += sc->nr_scanned;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002667 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Johannes Weiner0b064962014-08-06 16:06:12 -07002668 break;
2669
2670 if (sc->compaction_ready)
2671 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002672
2673 /*
Minchan Kim0e50ce32013-02-22 16:35:37 -08002674 * If we're getting trouble reclaiming, start doing
2675 * writepage even in laptop mode.
2676 */
2677 if (sc->priority < DEF_PRIORITY - 2)
2678 sc->may_writepage = 1;
2679
2680 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002681 * Try to write back as many pages as we just scanned. This
2682 * tends to cause slow streaming writers to write data to the
2683 * disk smoothly, at the dirtying rate, which is nice. But
2684 * that's undesirable in laptop mode, where we *want* lumpy
2685 * writeout. So in laptop mode, write out the whole world.
2686 */
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002687 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2688 if (total_scanned > writeback_threshold) {
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002689 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2690 WB_REASON_TRY_TO_FREE_PAGES);
Balbir Singh66e17072008-02-07 00:13:56 -08002691 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002692 }
Johannes Weiner0b064962014-08-06 16:06:12 -07002693 } while (--sc->priority >= 0);
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002694
Keika Kobayashi873b4772008-07-25 01:48:52 -07002695 delayacct_freepages_end();
2696
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002697 if (sc->nr_reclaimed)
2698 return sc->nr_reclaimed;
2699
Mel Gorman0cee34f2012-01-12 17:19:49 -08002700 /* Aborted reclaim to try compaction? don't OOM, then */
Johannes Weiner0b064962014-08-06 16:06:12 -07002701 if (sc->compaction_ready)
Mel Gorman73350842012-01-12 17:19:33 -08002702 return 1;
2703
Johannes Weiner241994ed2015-02-11 15:26:06 -08002704 /* Untapped cgroup reserves? Don't OOM, retry. */
2705 if (!sc->may_thrash) {
2706 sc->priority = initial_priority;
2707 sc->may_thrash = 1;
2708 goto retry;
2709 }
2710
Johannes Weiner2344d7e2014-08-06 16:06:15 -07002711 /* Any of the zones still reclaimable? Don't OOM. */
2712 if (zones_reclaimable)
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002713 return 1;
2714
2715 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002716}
2717
Mel Gorman55150612012-07-31 16:44:35 -07002718static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2719{
2720 struct zone *zone;
2721 unsigned long pfmemalloc_reserve = 0;
2722 unsigned long free_pages = 0;
2723 int i;
2724 bool wmark_ok;
2725
2726 for (i = 0; i <= ZONE_NORMAL; i++) {
2727 zone = &pgdat->node_zones[i];
Nishanth Aravamudanf012a842015-06-24 16:56:39 -07002728 if (!populated_zone(zone) ||
2729 zone_reclaimable_pages(zone) == 0)
Mel Gorman675becc2014-06-04 16:07:35 -07002730 continue;
2731
Mel Gorman55150612012-07-31 16:44:35 -07002732 pfmemalloc_reserve += min_wmark_pages(zone);
2733 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2734 }
2735
Mel Gorman675becc2014-06-04 16:07:35 -07002736 /* If there are no reserves (unexpected config) then do not throttle */
2737 if (!pfmemalloc_reserve)
2738 return true;
2739
Mel Gorman55150612012-07-31 16:44:35 -07002740 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2741
2742 /* kswapd must be awake if processes are being throttled */
2743 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2744 pgdat->classzone_idx = min(pgdat->classzone_idx,
2745 (enum zone_type)ZONE_NORMAL);
2746 wake_up_interruptible(&pgdat->kswapd_wait);
2747 }
2748
2749 return wmark_ok;
2750}
2751
2752/*
2753 * Throttle direct reclaimers if backing storage is backed by the network
2754 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2755 * depleted. kswapd will continue to make progress and wake the processes
Mel Gorman50694c22012-11-26 16:29:48 -08002756 * when the low watermark is reached.
2757 *
2758 * Returns true if a fatal signal was delivered during throttling. If this
2759 * happens, the page allocator should not consider triggering the OOM killer.
Mel Gorman55150612012-07-31 16:44:35 -07002760 */
Mel Gorman50694c22012-11-26 16:29:48 -08002761static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
Mel Gorman55150612012-07-31 16:44:35 -07002762 nodemask_t *nodemask)
2763{
Mel Gorman675becc2014-06-04 16:07:35 -07002764 struct zoneref *z;
Mel Gorman55150612012-07-31 16:44:35 -07002765 struct zone *zone;
Mel Gorman675becc2014-06-04 16:07:35 -07002766 pg_data_t *pgdat = NULL;
Mel Gorman55150612012-07-31 16:44:35 -07002767
2768 /*
2769 * Kernel threads should not be throttled as they may be indirectly
2770 * responsible for cleaning pages necessary for reclaim to make forward
2771 * progress. kjournald for example may enter direct reclaim while
2772 * committing a transaction where throttling it could forcing other
2773 * processes to block on log_wait_commit().
2774 */
2775 if (current->flags & PF_KTHREAD)
Mel Gorman50694c22012-11-26 16:29:48 -08002776 goto out;
2777
2778 /*
2779 * If a fatal signal is pending, this process should not throttle.
2780 * It should return quickly so it can exit and free its memory
2781 */
2782 if (fatal_signal_pending(current))
2783 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002784
Mel Gorman675becc2014-06-04 16:07:35 -07002785 /*
2786 * Check if the pfmemalloc reserves are ok by finding the first node
2787 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2788 * GFP_KERNEL will be required for allocating network buffers when
2789 * swapping over the network so ZONE_HIGHMEM is unusable.
2790 *
2791 * Throttling is based on the first usable node and throttled processes
2792 * wait on a queue until kswapd makes progress and wakes them. There
2793 * is an affinity then between processes waking up and where reclaim
2794 * progress has been made assuming the process wakes on the same node.
2795 * More importantly, processes running on remote nodes will not compete
2796 * for remote pfmemalloc reserves and processes on different nodes
2797 * should make reasonable progress.
2798 */
2799 for_each_zone_zonelist_nodemask(zone, z, zonelist,
Michael S. Tsirkin17636fa2015-01-26 12:58:41 -08002800 gfp_zone(gfp_mask), nodemask) {
Mel Gorman675becc2014-06-04 16:07:35 -07002801 if (zone_idx(zone) > ZONE_NORMAL)
2802 continue;
2803
2804 /* Throttle based on the first usable node */
2805 pgdat = zone->zone_pgdat;
2806 if (pfmemalloc_watermark_ok(pgdat))
2807 goto out;
2808 break;
2809 }
2810
2811 /* If no zone was usable by the allocation flags then do not throttle */
2812 if (!pgdat)
Mel Gorman50694c22012-11-26 16:29:48 -08002813 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002814
Mel Gorman68243e72012-07-31 16:44:39 -07002815 /* Account for the throttling */
2816 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2817
Mel Gorman55150612012-07-31 16:44:35 -07002818 /*
2819 * If the caller cannot enter the filesystem, it's possible that it
2820 * is due to the caller holding an FS lock or performing a journal
2821 * transaction in the case of a filesystem like ext[3|4]. In this case,
2822 * it is not safe to block on pfmemalloc_wait as kswapd could be
2823 * blocked waiting on the same lock. Instead, throttle for up to a
2824 * second before continuing.
2825 */
2826 if (!(gfp_mask & __GFP_FS)) {
2827 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2828 pfmemalloc_watermark_ok(pgdat), HZ);
Mel Gorman50694c22012-11-26 16:29:48 -08002829
2830 goto check_pending;
Mel Gorman55150612012-07-31 16:44:35 -07002831 }
2832
2833 /* Throttle until kswapd wakes the process */
2834 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2835 pfmemalloc_watermark_ok(pgdat));
Mel Gorman50694c22012-11-26 16:29:48 -08002836
2837check_pending:
2838 if (fatal_signal_pending(current))
2839 return true;
2840
2841out:
2842 return false;
Mel Gorman55150612012-07-31 16:44:35 -07002843}
2844
Mel Gormandac1d272008-04-28 02:12:12 -07002845unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002846 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08002847{
Mel Gorman33906bc2010-08-09 17:19:16 -07002848 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002849 struct scan_control sc = {
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002850 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weineree814fe2014-08-06 16:06:19 -07002851 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
2852 .order = order,
2853 .nodemask = nodemask,
2854 .priority = DEF_PRIORITY,
2855 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002856 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002857 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08002858 };
2859
Mel Gorman55150612012-07-31 16:44:35 -07002860 /*
Mel Gorman50694c22012-11-26 16:29:48 -08002861 * Do not enter reclaim if fatal signal was delivered while throttled.
2862 * 1 is returned so that the page allocator does not OOM kill at this
2863 * point.
Mel Gorman55150612012-07-31 16:44:35 -07002864 */
Mel Gorman50694c22012-11-26 16:29:48 -08002865 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
Mel Gorman55150612012-07-31 16:44:35 -07002866 return 1;
2867
Mel Gorman33906bc2010-08-09 17:19:16 -07002868 trace_mm_vmscan_direct_reclaim_begin(order,
2869 sc.may_writepage,
2870 gfp_mask);
2871
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002872 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Mel Gorman33906bc2010-08-09 17:19:16 -07002873
2874 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2875
2876 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002877}
2878
Andrew Mortonc255a452012-07-31 16:43:02 -07002879#ifdef CONFIG_MEMCG
Balbir Singh66e17072008-02-07 00:13:56 -08002880
Johannes Weiner72835c82012-01-12 17:18:32 -08002881unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002882 gfp_t gfp_mask, bool noswap,
Ying Han0ae5e892011-05-26 16:25:25 -07002883 struct zone *zone,
2884 unsigned long *nr_scanned)
Balbir Singh4e416952009-09-23 15:56:39 -07002885{
2886 struct scan_control sc = {
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07002887 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weineree814fe2014-08-06 16:06:19 -07002888 .target_mem_cgroup = memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002889 .may_writepage = !laptop_mode,
2890 .may_unmap = 1,
2891 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07002892 };
Johannes Weiner6b4f7792014-12-12 16:56:13 -08002893 unsigned long lru_pages;
Ying Han0ae5e892011-05-26 16:25:25 -07002894
Balbir Singh4e416952009-09-23 15:56:39 -07002895 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2896 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002897
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002898 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002899 sc.may_writepage,
2900 sc.gfp_mask);
2901
Balbir Singh4e416952009-09-23 15:56:39 -07002902 /*
2903 * NOTE: Although we can get the priority field, using it
2904 * here is not a good idea, since it limits the pages we can scan.
2905 * if we don't reclaim here, the shrink_zone from balance_pgdat
2906 * will pick up pages from other mem cgroup's as well. We hack
2907 * the priority and make it zero.
2908 */
Vladimir Davydov33377672016-01-20 15:02:59 -08002909 shrink_zone_memcg(zone, memcg, &sc, &lru_pages);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002910
2911 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2912
Ying Han0ae5e892011-05-26 16:25:25 -07002913 *nr_scanned = sc.nr_scanned;
Balbir Singh4e416952009-09-23 15:56:39 -07002914 return sc.nr_reclaimed;
2915}
2916
Johannes Weiner72835c82012-01-12 17:18:32 -08002917unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07002918 unsigned long nr_pages,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08002919 gfp_t gfp_mask,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07002920 bool may_swap)
Balbir Singh66e17072008-02-07 00:13:56 -08002921{
Balbir Singh4e416952009-09-23 15:56:39 -07002922 struct zonelist *zonelist;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002923 unsigned long nr_reclaimed;
Ying Han889976d2011-05-26 16:25:33 -07002924 int nid;
Balbir Singh66e17072008-02-07 00:13:56 -08002925 struct scan_control sc = {
Johannes Weinerb70a2a22014-10-09 15:28:56 -07002926 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Johannes Weineree814fe2014-08-06 16:06:19 -07002927 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2928 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2929 .target_mem_cgroup = memcg,
2930 .priority = DEF_PRIORITY,
Balbir Singh66e17072008-02-07 00:13:56 -08002931 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002932 .may_unmap = 1,
Johannes Weinerb70a2a22014-10-09 15:28:56 -07002933 .may_swap = may_swap,
Ying Hana09ed5e2011-05-24 17:12:26 -07002934 };
Balbir Singh66e17072008-02-07 00:13:56 -08002935
Ying Han889976d2011-05-26 16:25:33 -07002936 /*
2937 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2938 * take care of from where we get pages. So the node where we start the
2939 * scan does not need to be the current node.
2940 */
Johannes Weiner72835c82012-01-12 17:18:32 -08002941 nid = mem_cgroup_select_victim_node(memcg);
Ying Han889976d2011-05-26 16:25:33 -07002942
2943 zonelist = NODE_DATA(nid)->node_zonelists;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002944
2945 trace_mm_vmscan_memcg_reclaim_begin(0,
2946 sc.may_writepage,
2947 sc.gfp_mask);
2948
Vladimir Davydov3115cd92014-04-03 14:47:22 -07002949 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002950
2951 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2952
2953 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002954}
2955#endif
2956
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002957static void age_active_anon(struct zone *zone, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002958{
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002959 struct mem_cgroup *memcg;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002960
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002961 if (!total_swap_pages)
2962 return;
2963
2964 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2965 do {
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07002966 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002967
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07002968 if (inactive_anon_is_low(lruvec))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002969 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002970 sc, LRU_ACTIVE_ANON);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002971
2972 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2973 } while (memcg);
Johannes Weinerf16015f2012-01-12 17:17:52 -08002974}
2975
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07002976static bool zone_balanced(struct zone *zone, int order, bool highorder,
2977 unsigned long balance_gap, int classzone_idx)
Johannes Weiner60cefed2012-11-29 13:54:23 -08002978{
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07002979 unsigned long mark = high_wmark_pages(zone) + balance_gap;
Johannes Weiner60cefed2012-11-29 13:54:23 -08002980
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07002981 /*
2982 * When checking from pgdat_balanced(), kswapd should stop and sleep
2983 * when it reaches the high order-0 watermark and let kcompactd take
2984 * over. Other callers such as wakeup_kswapd() want to determine the
2985 * true high-order watermark.
2986 */
2987 if (IS_ENABLED(CONFIG_COMPACTION) && !highorder) {
2988 mark += (1UL << order);
2989 order = 0;
2990 }
Johannes Weiner60cefed2012-11-29 13:54:23 -08002991
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07002992 return zone_watermark_ok_safe(zone, order, mark, classzone_idx);
Johannes Weiner60cefed2012-11-29 13:54:23 -08002993}
2994
Mel Gorman1741c872011-01-13 15:46:21 -08002995/*
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002996 * pgdat_balanced() is used when checking if a node is balanced.
2997 *
2998 * For order-0, all zones must be balanced!
2999 *
3000 * For high-order allocations only zones that meet watermarks and are in a
3001 * zone allowed by the callers classzone_idx are added to balanced_pages. The
3002 * total of balanced pages must be at least 25% of the zones allowed by
3003 * classzone_idx for the node to be considered balanced. Forcing all zones to
3004 * be balanced for high orders can cause excessive reclaim when there are
3005 * imbalanced zones.
Mel Gorman1741c872011-01-13 15:46:21 -08003006 * The choice of 25% is due to
3007 * o a 16M DMA zone that is balanced will not balance a zone on any
3008 * reasonable sized machine
3009 * o On all other machines, the top zone must be at least a reasonable
Lucas De Marchi25985ed2011-03-30 22:57:33 -03003010 * percentage of the middle zones. For example, on 32-bit x86, highmem
Mel Gorman1741c872011-01-13 15:46:21 -08003011 * would need to be at least 256M for it to be balance a whole node.
3012 * Similarly, on x86-64 the Normal zone would need to be at least 1G
3013 * to balance a node on its own. These seemed like reasonable ratios.
3014 */
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01003015static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
Mel Gorman1741c872011-01-13 15:46:21 -08003016{
Jiang Liub40da042013-02-22 16:33:52 -08003017 unsigned long managed_pages = 0;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01003018 unsigned long balanced_pages = 0;
Mel Gorman1741c872011-01-13 15:46:21 -08003019 int i;
3020
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01003021 /* Check the watermark levels */
3022 for (i = 0; i <= classzone_idx; i++) {
3023 struct zone *zone = pgdat->node_zones + i;
Mel Gorman1741c872011-01-13 15:46:21 -08003024
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01003025 if (!populated_zone(zone))
3026 continue;
3027
Jiang Liub40da042013-02-22 16:33:52 -08003028 managed_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01003029
3030 /*
3031 * A special case here:
3032 *
3033 * balance_pgdat() skips over all_unreclaimable after
3034 * DEF_PRIORITY. Effectively, it considers them balanced so
3035 * they must be considered balanced here as well!
3036 */
Lisa Du6e543d52013-09-11 14:22:36 -07003037 if (!zone_reclaimable(zone)) {
Jiang Liub40da042013-02-22 16:33:52 -08003038 balanced_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01003039 continue;
3040 }
3041
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003042 if (zone_balanced(zone, order, false, 0, i))
Jiang Liub40da042013-02-22 16:33:52 -08003043 balanced_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01003044 else if (!order)
3045 return false;
3046 }
3047
3048 if (order)
Jiang Liub40da042013-02-22 16:33:52 -08003049 return balanced_pages >= (managed_pages >> 2);
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01003050 else
3051 return true;
Mel Gorman1741c872011-01-13 15:46:21 -08003052}
3053
Mel Gorman55150612012-07-31 16:44:35 -07003054/*
3055 * Prepare kswapd for sleeping. This verifies that there are no processes
3056 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3057 *
3058 * Returns true if kswapd is ready to sleep
3059 */
3060static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
Mel Gormandc83edd2011-01-13 15:46:26 -08003061 int classzone_idx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08003062{
Mel Gormanf50de2d2009-12-14 17:58:53 -08003063 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
3064 if (remaining)
Mel Gorman55150612012-07-31 16:44:35 -07003065 return false;
3066
3067 /*
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08003068 * The throttled processes are normally woken up in balance_pgdat() as
3069 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
3070 * race between when kswapd checks the watermarks and a process gets
3071 * throttled. There is also a potential race if processes get
3072 * throttled, kswapd wakes, a large process exits thereby balancing the
3073 * zones, which causes kswapd to exit balance_pgdat() before reaching
3074 * the wake up checks. If kswapd is going to sleep, no process should
3075 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3076 * the wake up is premature, processes will wake kswapd and get
3077 * throttled again. The difference from wake ups in balance_pgdat() is
3078 * that here we are under prepare_to_wait().
Mel Gorman55150612012-07-31 16:44:35 -07003079 */
Vlastimil Babka9e5e3662015-01-08 14:32:40 -08003080 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3081 wake_up_all(&pgdat->pfmemalloc_wait);
Mel Gormanf50de2d2009-12-14 17:58:53 -08003082
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01003083 return pgdat_balanced(pgdat, order, classzone_idx);
Mel Gormanf50de2d2009-12-14 17:58:53 -08003084}
3085
Linus Torvalds1da177e2005-04-16 15:20:36 -07003086/*
Mel Gorman75485362013-07-03 15:01:42 -07003087 * kswapd shrinks the zone by the number of pages required to reach
3088 * the high watermark.
Mel Gormanb8e83b92013-07-03 15:01:45 -07003089 *
3090 * Returns true if kswapd scanned at least the requested number of pages to
Mel Gorman283aba92013-07-03 15:01:51 -07003091 * reclaim or if the lack of progress was due to pages under writeback.
3092 * This is used to determine if the scanning priority needs to be raised.
Mel Gorman75485362013-07-03 15:01:42 -07003093 */
Mel Gormanb8e83b92013-07-03 15:01:45 -07003094static bool kswapd_shrink_zone(struct zone *zone,
Mel Gorman7c954f62013-07-03 15:01:54 -07003095 int classzone_idx,
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003096 struct scan_control *sc)
Mel Gorman75485362013-07-03 15:01:42 -07003097{
Mel Gorman7c954f62013-07-03 15:01:54 -07003098 unsigned long balance_gap;
Mel Gorman7c954f62013-07-03 15:01:54 -07003099 bool lowmem_pressure;
Mel Gorman75485362013-07-03 15:01:42 -07003100
3101 /* Reclaim above the high watermark. */
3102 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
Mel Gorman7c954f62013-07-03 15:01:54 -07003103
3104 /*
Mel Gorman7c954f62013-07-03 15:01:54 -07003105 * We put equal pressure on every zone, unless one zone has way too
3106 * many pages free already. The "too many pages" is defined as the
3107 * high wmark plus a "gap" where the gap is either the low
3108 * watermark or 1% of the zone, whichever is smaller.
3109 */
Jianyu Zhan4be89a32014-06-04 16:10:38 -07003110 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
3111 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
Mel Gorman7c954f62013-07-03 15:01:54 -07003112
3113 /*
3114 * If there is no low memory pressure or the zone is balanced then no
3115 * reclaim is necessary
3116 */
3117 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003118 if (!lowmem_pressure && zone_balanced(zone, sc->order, false,
Mel Gorman7c954f62013-07-03 15:01:54 -07003119 balance_gap, classzone_idx))
3120 return true;
3121
Johannes Weiner6b4f7792014-12-12 16:56:13 -08003122 shrink_zone(zone, sc, zone_idx(zone) == classzone_idx);
Mel Gorman75485362013-07-03 15:01:42 -07003123
Johannes Weiner57054652014-10-09 15:28:17 -07003124 clear_bit(ZONE_WRITEBACK, &zone->flags);
Mel Gorman283aba92013-07-03 15:01:51 -07003125
Mel Gorman7c954f62013-07-03 15:01:54 -07003126 /*
3127 * If a zone reaches its high watermark, consider it to be no longer
3128 * congested. It's possible there are dirty pages backed by congested
3129 * BDIs but as pressure is relieved, speculatively avoid congestion
3130 * waits.
3131 */
Lisa Du6e543d52013-09-11 14:22:36 -07003132 if (zone_reclaimable(zone) &&
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003133 zone_balanced(zone, sc->order, false, 0, classzone_idx)) {
Johannes Weiner57054652014-10-09 15:28:17 -07003134 clear_bit(ZONE_CONGESTED, &zone->flags);
3135 clear_bit(ZONE_DIRTY, &zone->flags);
Mel Gorman7c954f62013-07-03 15:01:54 -07003136 }
3137
Mel Gormanb8e83b92013-07-03 15:01:45 -07003138 return sc->nr_scanned >= sc->nr_to_reclaim;
Mel Gorman75485362013-07-03 15:01:42 -07003139}
3140
3141/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003142 * For kswapd, balance_pgdat() will work across all this node's zones until
Mel Gorman41858962009-06-16 15:32:12 -07003143 * they are all at high_wmark_pages(zone).
Linus Torvalds1da177e2005-04-16 15:20:36 -07003144 *
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003145 * Returns the highest zone idx kswapd was reclaiming at
Linus Torvalds1da177e2005-04-16 15:20:36 -07003146 *
3147 * There is special handling here for zones which are full of pinned pages.
3148 * This can happen if the pages are all mlocked, or if they are all used by
3149 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
3150 * What we do is to detect the case where all pages in the zone have been
3151 * scanned twice and there has been zero successful reclaim. Mark the zone as
3152 * dead and from now on, only perform a short scan. Basically we're polling
3153 * the zone for when the problem goes away.
3154 *
3155 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07003156 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3157 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
3158 * lower zones regardless of the number of free pages in the lower zones. This
3159 * interoperates with the page allocator fallback scheme to ensure that aging
3160 * of pages is balanced across the zones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003161 */
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003162static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003163{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003164 int i;
Mel Gorman99504742011-01-13 15:46:20 -08003165 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
Andrew Morton0608f432013-09-24 15:27:41 -07003166 unsigned long nr_soft_reclaimed;
3167 unsigned long nr_soft_scanned;
Andrew Morton179e9632006-03-22 00:08:18 -08003168 struct scan_control sc = {
3169 .gfp_mask = GFP_KERNEL,
Johannes Weineree814fe2014-08-06 16:06:19 -07003170 .order = order,
Mel Gormanb8e83b92013-07-03 15:01:45 -07003171 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003172 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003173 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003174 .may_swap = 1,
Andrew Morton179e9632006-03-22 00:08:18 -08003175 };
Christoph Lameterf8891e52006-06-30 01:55:45 -07003176 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003177
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003178 do {
Mel Gormanb8e83b92013-07-03 15:01:45 -07003179 bool raise_priority = true;
3180
3181 sc.nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003182
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003183 /*
3184 * Scan in the highmem->dma direction for the highest
3185 * zone which needs scanning
3186 */
3187 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
3188 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003189
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003190 if (!populated_zone(zone))
3191 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003192
Lisa Du6e543d52013-09-11 14:22:36 -07003193 if (sc.priority != DEF_PRIORITY &&
3194 !zone_reclaimable(zone))
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003195 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003196
Rik van Riel556adec2008-10-18 20:26:34 -07003197 /*
3198 * Do some background aging of the anon list, to give
3199 * pages a chance to be referenced before reclaiming.
3200 */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003201 age_active_anon(zone, &sc);
Rik van Riel556adec2008-10-18 20:26:34 -07003202
Mel Gormancc715d92012-03-21 16:34:00 -07003203 /*
3204 * If the number of buffer_heads in the machine
3205 * exceeds the maximum allowed level and this node
3206 * has a highmem zone, force kswapd to reclaim from
3207 * it to relieve lowmem pressure.
3208 */
3209 if (buffer_heads_over_limit && is_highmem_idx(i)) {
3210 end_zone = i;
3211 break;
3212 }
3213
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003214 if (!zone_balanced(zone, order, false, 0, 0)) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003215 end_zone = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003216 break;
Shaohua Li439423f2011-08-25 15:59:12 -07003217 } else {
Mel Gormand43006d2013-07-03 15:01:50 -07003218 /*
3219 * If balanced, clear the dirty and congested
3220 * flags
3221 */
Johannes Weiner57054652014-10-09 15:28:17 -07003222 clear_bit(ZONE_CONGESTED, &zone->flags);
3223 clear_bit(ZONE_DIRTY, &zone->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003224 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003225 }
Zlatko Calusicdafcb732013-02-22 16:32:34 -08003226
Mel Gormanb8e83b92013-07-03 15:01:45 -07003227 if (i < 0)
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003228 goto out;
3229
Linus Torvalds1da177e2005-04-16 15:20:36 -07003230 /*
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003231 * If we're getting trouble reclaiming, start doing writepage
3232 * even in laptop mode.
3233 */
3234 if (sc.priority < DEF_PRIORITY - 2)
3235 sc.may_writepage = 1;
3236
3237 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003238 * Now scan the zone in the dma->highmem direction, stopping
3239 * at the last zone which needs scanning.
3240 *
3241 * We do this because the page allocator works in the opposite
3242 * direction. This prevents the page allocator from allocating
3243 * pages behind kswapd's direction of progress, which would
3244 * cause too much scanning of the lower zones.
3245 */
3246 for (i = 0; i <= end_zone; i++) {
3247 struct zone *zone = pgdat->node_zones + i;
3248
Con Kolivasf3fe6512006-01-06 00:11:15 -08003249 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003250 continue;
3251
Lisa Du6e543d52013-09-11 14:22:36 -07003252 if (sc.priority != DEF_PRIORITY &&
3253 !zone_reclaimable(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003254 continue;
3255
Linus Torvalds1da177e2005-04-16 15:20:36 -07003256 sc.nr_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07003257
Andrew Morton0608f432013-09-24 15:27:41 -07003258 nr_soft_scanned = 0;
3259 /*
3260 * Call soft limit reclaim before calling shrink_zone.
3261 */
3262 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
3263 order, sc.gfp_mask,
3264 &nr_soft_scanned);
3265 sc.nr_reclaimed += nr_soft_reclaimed;
3266
Rik van Riel32a43302007-10-16 01:24:50 -07003267 /*
Mel Gorman7c954f62013-07-03 15:01:54 -07003268 * There should be no need to raise the scanning
3269 * priority if enough pages are already being scanned
3270 * that that high watermark would be met at 100%
3271 * efficiency.
Rik van Riel32a43302007-10-16 01:24:50 -07003272 */
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003273 if (kswapd_shrink_zone(zone, end_zone, &sc))
Mel Gorman7c954f62013-07-03 15:01:54 -07003274 raise_priority = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003275 }
Mel Gorman55150612012-07-31 16:44:35 -07003276
3277 /*
3278 * If the low watermark is met there is no need for processes
3279 * to be throttled on pfmemalloc_wait as they should not be
3280 * able to safely make forward progress. Wake them
3281 */
3282 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3283 pfmemalloc_watermark_ok(pgdat))
Vlastimil Babkacfc51152015-02-11 15:25:12 -08003284 wake_up_all(&pgdat->pfmemalloc_wait);
Mel Gorman55150612012-07-31 16:44:35 -07003285
Mel Gormanb8e83b92013-07-03 15:01:45 -07003286 /* Check if kswapd should be suspending */
3287 if (try_to_freeze() || kthread_should_stop())
3288 break;
3289
3290 /*
3291 * Raise priority if scanning rate is too low or there was no
3292 * progress in reclaiming pages
3293 */
3294 if (raise_priority || !sc.nr_reclaimed)
3295 sc.priority--;
Mel Gorman9aa41342013-07-03 15:01:48 -07003296 } while (sc.priority >= 1 &&
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003297 !pgdat_balanced(pgdat, order, classzone_idx));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003298
Mel Gormanb8e83b92013-07-03 15:01:45 -07003299out:
Mel Gorman0abdee22011-01-13 15:46:22 -08003300 /*
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003301 * Return the highest zone idx we were reclaiming at so
3302 * prepare_kswapd_sleep() makes the same decisions as here.
Mel Gorman0abdee22011-01-13 15:46:22 -08003303 */
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003304 return end_zone;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003305}
3306
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003307static void kswapd_try_to_sleep(pg_data_t *pgdat, int order,
3308 int classzone_idx, int balanced_classzone_idx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003309{
3310 long remaining = 0;
3311 DEFINE_WAIT(wait);
3312
3313 if (freezing(current) || kthread_should_stop())
3314 return;
3315
3316 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3317
3318 /* Try to sleep for a short interval */
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003319 if (prepare_kswapd_sleep(pgdat, order, remaining,
3320 balanced_classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003321 remaining = schedule_timeout(HZ/10);
3322 finish_wait(&pgdat->kswapd_wait, &wait);
3323 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3324 }
3325
3326 /*
3327 * After a short sleep, check if it was a premature sleep. If not, then
3328 * go fully to sleep until explicitly woken up.
3329 */
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003330 if (prepare_kswapd_sleep(pgdat, order, remaining,
3331 balanced_classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003332 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3333
3334 /*
3335 * vmstat counters are not perfectly accurate and the estimated
3336 * value for counters such as NR_FREE_PAGES can deviate from the
3337 * true value by nr_online_cpus * threshold. To avoid the zone
3338 * watermarks being breached while under pressure, we reduce the
3339 * per-cpu vmstat threshold while kswapd is awake and restore
3340 * them before going back to sleep.
3341 */
3342 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003343
Mel Gorman62997022012-10-08 16:32:47 -07003344 /*
3345 * Compaction records what page blocks it recently failed to
3346 * isolate pages from and skips them in the future scanning.
3347 * When kswapd is going to sleep, it is reasonable to assume
3348 * that pages and compaction may succeed so reset the cache.
3349 */
3350 reset_isolation_suitable(pgdat);
3351
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003352 /*
3353 * We have freed the memory, now we should compact it to make
3354 * allocation of the requested order possible.
3355 */
3356 wakeup_kcompactd(pgdat, order, classzone_idx);
3357
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003358 if (!kthread_should_stop())
3359 schedule();
3360
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003361 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3362 } else {
3363 if (remaining)
3364 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3365 else
3366 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3367 }
3368 finish_wait(&pgdat->kswapd_wait, &wait);
3369}
3370
Linus Torvalds1da177e2005-04-16 15:20:36 -07003371/*
3372 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003373 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003374 *
3375 * This basically trickles out pages so that we have _some_
3376 * free memory available even if there is no other activity
3377 * that frees anything up. This is needed for things like routing
3378 * etc, where we otherwise might have all activity going on in
3379 * asynchronous contexts that cannot page things out.
3380 *
3381 * If there are applications that are active memory-allocators
3382 * (most normal use), this basically shouldn't matter.
3383 */
3384static int kswapd(void *p)
3385{
Mel Gorman215ddd62011-07-08 15:39:40 -07003386 unsigned long order, new_order;
3387 int classzone_idx, new_classzone_idx;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003388 int balanced_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003389 pg_data_t *pgdat = (pg_data_t*)p;
3390 struct task_struct *tsk = current;
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003391
Linus Torvalds1da177e2005-04-16 15:20:36 -07003392 struct reclaim_state reclaim_state = {
3393 .reclaimed_slab = 0,
3394 };
Rusty Russella70f7302009-03-13 14:49:46 +10303395 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003396
Nick Piggincf40bd12009-01-21 08:12:39 +01003397 lockdep_set_current_reclaim_state(GFP_KERNEL);
3398
Rusty Russell174596a2009-01-01 10:12:29 +10303399 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07003400 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003401 current->reclaim_state = &reclaim_state;
3402
3403 /*
3404 * Tell the memory management that we're a "memory allocator",
3405 * and that if we need more memory we should get access to it
3406 * regardless (see "__alloc_pages()"). "kswapd" should
3407 * never get caught in the normal page freeing logic.
3408 *
3409 * (Kswapd normally doesn't need memory anyway, but sometimes
3410 * you need a small amount of memory in order to be able to
3411 * page out something else, and this flag essentially protects
3412 * us from recursively trying to free more memory as we're
3413 * trying to free the first piece of memory in the first place).
3414 */
Christoph Lameter930d9152006-01-08 01:00:47 -08003415 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07003416 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003417
Mel Gorman215ddd62011-07-08 15:39:40 -07003418 order = new_order = 0;
3419 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003420 balanced_classzone_idx = classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003421 for ( ; ; ) {
Jeff Liu6f6313d2012-12-11 16:02:48 -08003422 bool ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07003423
Mel Gorman215ddd62011-07-08 15:39:40 -07003424 /*
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003425 * While we were reclaiming, there might have been another
3426 * wakeup, so check the values.
Mel Gorman215ddd62011-07-08 15:39:40 -07003427 */
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003428 new_order = pgdat->kswapd_max_order;
3429 new_classzone_idx = pgdat->classzone_idx;
3430 pgdat->kswapd_max_order = 0;
3431 pgdat->classzone_idx = pgdat->nr_zones - 1;
Mel Gorman215ddd62011-07-08 15:39:40 -07003432
Mel Gorman99504742011-01-13 15:46:20 -08003433 if (order < new_order || classzone_idx > new_classzone_idx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003434 /*
3435 * Don't sleep if someone wants a larger 'order'
Mel Gorman99504742011-01-13 15:46:20 -08003436 * allocation or has tigher zone constraints
Linus Torvalds1da177e2005-04-16 15:20:36 -07003437 */
3438 order = new_order;
Mel Gorman99504742011-01-13 15:46:20 -08003439 classzone_idx = new_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003440 } else {
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003441 kswapd_try_to_sleep(pgdat, order, classzone_idx,
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003442 balanced_classzone_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003443 order = pgdat->kswapd_max_order;
Mel Gorman99504742011-01-13 15:46:20 -08003444 classzone_idx = pgdat->classzone_idx;
Alex,Shif0dfcde2011-10-31 17:08:45 -07003445 new_order = order;
3446 new_classzone_idx = classzone_idx;
Mel Gorman4d405022011-01-13 15:46:23 -08003447 pgdat->kswapd_max_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07003448 pgdat->classzone_idx = pgdat->nr_zones - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003449 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003450
David Rientjes8fe23e02009-12-14 17:58:33 -08003451 ret = try_to_freeze();
3452 if (kthread_should_stop())
3453 break;
3454
3455 /*
3456 * We can speed up thawing tasks if we don't call balance_pgdat
3457 * after returning from the refrigerator
3458 */
Mel Gorman33906bc2010-08-09 17:19:16 -07003459 if (!ret) {
3460 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003461 balanced_classzone_idx = balance_pgdat(pgdat, order,
3462 classzone_idx);
Mel Gorman33906bc2010-08-09 17:19:16 -07003463 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003464 }
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003465
Johannes Weiner71abdc12014-06-06 14:35:35 -07003466 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003467 current->reclaim_state = NULL;
Johannes Weiner71abdc12014-06-06 14:35:35 -07003468 lockdep_clear_current_reclaim_state();
3469
Linus Torvalds1da177e2005-04-16 15:20:36 -07003470 return 0;
3471}
3472
3473/*
3474 * A zone is low on free memory, so wake its kswapd task to service it.
3475 */
Mel Gorman99504742011-01-13 15:46:20 -08003476void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003477{
3478 pg_data_t *pgdat;
3479
Con Kolivasf3fe6512006-01-06 00:11:15 -08003480 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003481 return;
3482
Vladimir Davydov344736f2014-10-20 15:50:30 +04003483 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003484 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08003485 pgdat = zone->zone_pgdat;
Mel Gorman99504742011-01-13 15:46:20 -08003486 if (pgdat->kswapd_max_order < order) {
Mel Gorman88f5acf2011-01-13 15:45:41 -08003487 pgdat->kswapd_max_order = order;
Mel Gorman99504742011-01-13 15:46:20 -08003488 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3489 }
Con Kolivas8d0986e2005-09-13 01:25:07 -07003490 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003491 return;
Vlastimil Babkaaccf6242016-03-17 14:18:15 -07003492 if (zone_balanced(zone, order, true, 0, 0))
Mel Gorman88f5acf2011-01-13 15:45:41 -08003493 return;
3494
3495 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07003496 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003497}
3498
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003499#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07003500/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003501 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003502 * freed pages.
3503 *
3504 * Rather than trying to age LRUs the aim is to preserve the overall
3505 * LRU order by reclaiming preferentially
3506 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07003507 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003508unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003509{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003510 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003511 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003512 .nr_to_reclaim = nr_to_reclaim,
Johannes Weineree814fe2014-08-06 16:06:19 -07003513 .gfp_mask = GFP_HIGHUSER_MOVABLE,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003514 .priority = DEF_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003515 .may_writepage = 1,
3516 .may_unmap = 1,
3517 .may_swap = 1,
3518 .hibernation_mode = 1,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003519 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003520 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003521 struct task_struct *p = current;
3522 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003523
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003524 p->flags |= PF_MEMALLOC;
3525 lockdep_set_current_reclaim_state(sc.gfp_mask);
3526 reclaim_state.reclaimed_slab = 0;
3527 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08003528
Vladimir Davydov3115cd92014-04-03 14:47:22 -07003529 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003530
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003531 p->reclaim_state = NULL;
3532 lockdep_clear_current_reclaim_state();
3533 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003534
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003535 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003536}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003537#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003538
Linus Torvalds1da177e2005-04-16 15:20:36 -07003539/* It's optimal to keep kswapds on the same CPUs as their memory, but
3540 not required for correctness. So if the last cpu in a node goes
3541 away, we get changed to run anywhere: as the first one comes back,
3542 restore their cpu bindings. */
Greg Kroah-Hartmanfcb35a92012-12-21 15:01:06 -08003543static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3544 void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003545{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07003546 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003547
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003548 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003549 for_each_node_state(nid, N_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07003550 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10303551 const struct cpumask *mask;
3552
3553 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07003554
Rusty Russell3e597942009-01-01 10:12:24 +10303555 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003556 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07003557 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003558 }
3559 }
3560 return NOTIFY_OK;
3561}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003562
Yasunori Goto3218ae12006-06-27 02:53:33 -07003563/*
3564 * This kswapd start function will be called by init and node-hot-add.
3565 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3566 */
3567int kswapd_run(int nid)
3568{
3569 pg_data_t *pgdat = NODE_DATA(nid);
3570 int ret = 0;
3571
3572 if (pgdat->kswapd)
3573 return 0;
3574
3575 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3576 if (IS_ERR(pgdat->kswapd)) {
3577 /* failure at boot is fatal */
3578 BUG_ON(system_state == SYSTEM_BOOTING);
Gavin Shand5dc0ad2012-10-08 16:29:27 -07003579 pr_err("Failed to start kswapd on node %d\n", nid);
3580 ret = PTR_ERR(pgdat->kswapd);
Xishi Qiud72515b2013-04-17 15:58:34 -07003581 pgdat->kswapd = NULL;
Yasunori Goto3218ae12006-06-27 02:53:33 -07003582 }
3583 return ret;
3584}
3585
David Rientjes8fe23e02009-12-14 17:58:33 -08003586/*
Jiang Liud8adde12012-07-11 14:01:52 -07003587 * Called by memory hotplug when all memory in a node is offlined. Caller must
Vladimir Davydovbfc8c902014-06-04 16:07:18 -07003588 * hold mem_hotplug_begin/end().
David Rientjes8fe23e02009-12-14 17:58:33 -08003589 */
3590void kswapd_stop(int nid)
3591{
3592 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3593
Jiang Liud8adde12012-07-11 14:01:52 -07003594 if (kswapd) {
David Rientjes8fe23e02009-12-14 17:58:33 -08003595 kthread_stop(kswapd);
Jiang Liud8adde12012-07-11 14:01:52 -07003596 NODE_DATA(nid)->kswapd = NULL;
3597 }
David Rientjes8fe23e02009-12-14 17:58:33 -08003598}
3599
Linus Torvalds1da177e2005-04-16 15:20:36 -07003600static int __init kswapd_init(void)
3601{
Yasunori Goto3218ae12006-06-27 02:53:33 -07003602 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08003603
Linus Torvalds1da177e2005-04-16 15:20:36 -07003604 swap_setup();
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003605 for_each_node_state(nid, N_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07003606 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003607 hotcpu_notifier(cpu_callback, 0);
3608 return 0;
3609}
3610
3611module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003612
3613#ifdef CONFIG_NUMA
3614/*
3615 * Zone reclaim mode
3616 *
3617 * If non-zero call zone_reclaim when the number of free pages falls below
3618 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08003619 */
3620int zone_reclaim_mode __read_mostly;
3621
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003622#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07003623#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003624#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003625#define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003626
Christoph Lameter9eeff232006-01-18 17:42:31 -08003627/*
Christoph Lametera92f7122006-02-01 03:05:32 -08003628 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3629 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3630 * a zone.
3631 */
3632#define ZONE_RECLAIM_PRIORITY 4
3633
Christoph Lameter9eeff232006-01-18 17:42:31 -08003634/*
Christoph Lameter96146342006-07-03 00:24:13 -07003635 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3636 * occur.
3637 */
3638int sysctl_min_unmapped_ratio = 1;
3639
3640/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003641 * If the number of slab pages in a zone grows beyond this percentage then
3642 * slab reclaim needs to occur.
3643 */
3644int sysctl_min_slab_ratio = 5;
3645
Mel Gorman90afa5d2009-06-16 15:33:20 -07003646static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3647{
3648 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3649 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3650 zone_page_state(zone, NR_ACTIVE_FILE);
3651
3652 /*
3653 * It's possible for there to be more file mapped pages than
3654 * accounted for by the pages on the file LRU lists because
3655 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3656 */
3657 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3658}
3659
3660/* Work out how many page cache pages we can reclaim in this reclaim_mode */
Alexandru Moised031a152015-11-05 18:48:08 -08003661static unsigned long zone_pagecache_reclaimable(struct zone *zone)
Mel Gorman90afa5d2009-06-16 15:33:20 -07003662{
Alexandru Moised031a152015-11-05 18:48:08 -08003663 unsigned long nr_pagecache_reclaimable;
3664 unsigned long delta = 0;
Mel Gorman90afa5d2009-06-16 15:33:20 -07003665
3666 /*
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003667 * If RECLAIM_UNMAP is set, then all file pages are considered
Mel Gorman90afa5d2009-06-16 15:33:20 -07003668 * potentially reclaimable. Otherwise, we have to worry about
3669 * pages like swapcache and zone_unmapped_file_pages() provides
3670 * a better estimate
3671 */
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003672 if (zone_reclaim_mode & RECLAIM_UNMAP)
Mel Gorman90afa5d2009-06-16 15:33:20 -07003673 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3674 else
3675 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3676
3677 /* If we can't clean pages, remove dirty pages from consideration */
3678 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3679 delta += zone_page_state(zone, NR_FILE_DIRTY);
3680
3681 /* Watch for any possible underflows due to delta */
3682 if (unlikely(delta > nr_pagecache_reclaimable))
3683 delta = nr_pagecache_reclaimable;
3684
3685 return nr_pagecache_reclaimable - delta;
3686}
3687
Christoph Lameter0ff38492006-09-25 23:31:52 -07003688/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08003689 * Try to free up some pages from this zone through reclaim.
3690 */
Andrew Morton179e9632006-03-22 00:08:18 -08003691static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003692{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003693 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08003694 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003695 struct task_struct *p = current;
3696 struct reclaim_state reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08003697 struct scan_control sc = {
Andrew Morton62b726c2013-02-22 16:32:24 -08003698 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Ming Lei21caf2f2013-02-22 16:34:08 -08003699 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
Johannes Weinerbd2f6192009-03-31 15:19:38 -07003700 .order = order,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003701 .priority = ZONE_RECLAIM_PRIORITY,
Johannes Weineree814fe2014-08-06 16:06:19 -07003702 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003703 .may_unmap = !!(zone_reclaim_mode & RECLAIM_UNMAP),
Johannes Weineree814fe2014-08-06 16:06:19 -07003704 .may_swap = 1,
Andrew Morton179e9632006-03-22 00:08:18 -08003705 };
Christoph Lameter9eeff232006-01-18 17:42:31 -08003706
Christoph Lameter9eeff232006-01-18 17:42:31 -08003707 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08003708 /*
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003709 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
Christoph Lameterd4f77962006-02-24 13:04:22 -08003710 * and we also need to be able to write out pages for RECLAIM_WRITE
Zhihui Zhang95bbc0c2015-06-24 16:56:42 -07003711 * and RECLAIM_UNMAP.
Christoph Lameterd4f77962006-02-24 13:04:22 -08003712 */
3713 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003714 lockdep_set_current_reclaim_state(gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08003715 reclaim_state.reclaimed_slab = 0;
3716 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08003717
Mel Gorman90afa5d2009-06-16 15:33:20 -07003718 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07003719 /*
3720 * Free memory by calling shrink zone with increasing
3721 * priorities until we have enough memory freed.
3722 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07003723 do {
Johannes Weiner6b4f7792014-12-12 16:56:13 -08003724 shrink_zone(zone, &sc, true);
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003725 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
Christoph Lameter0ff38492006-09-25 23:31:52 -07003726 }
Christoph Lameterc84db232006-02-01 03:05:29 -08003727
Christoph Lameter9eeff232006-01-18 17:42:31 -08003728 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08003729 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003730 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08003731 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003732}
Andrew Morton179e9632006-03-22 00:08:18 -08003733
3734int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3735{
Andrew Morton179e9632006-03-22 00:08:18 -08003736 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07003737 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003738
3739 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003740 * Zone reclaim reclaims unmapped file backed pages and
3741 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07003742 *
Christoph Lameter96146342006-07-03 00:24:13 -07003743 * A small portion of unmapped file backed pages is needed for
3744 * file I/O otherwise pages read by file I/O will be immediately
3745 * thrown out if the zone is overallocated. So we do not reclaim
3746 * if less than a specified percentage of the zone is used by
3747 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08003748 */
Mel Gorman90afa5d2009-06-16 15:33:20 -07003749 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3750 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003751 return ZONE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08003752
Lisa Du6e543d52013-09-11 14:22:36 -07003753 if (!zone_reclaimable(zone))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003754 return ZONE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07003755
Andrew Morton179e9632006-03-22 00:08:18 -08003756 /*
David Rientjesd773ed62007-10-16 23:26:01 -07003757 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08003758 */
Mel Gormand0164ad2015-11-06 16:28:21 -08003759 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003760 return ZONE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08003761
3762 /*
3763 * Only run zone reclaim on the local zone or on zones that do not
3764 * have associated processors. This will favor the local processor
3765 * over remote processors and spread off node memory allocations
3766 * as wide as possible.
3767 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07003768 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07003769 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Mel Gormanfa5e0842009-06-16 15:33:22 -07003770 return ZONE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07003771
Johannes Weiner57054652014-10-09 15:28:17 -07003772 if (test_and_set_bit(ZONE_RECLAIM_LOCKED, &zone->flags))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003773 return ZONE_RECLAIM_NOSCAN;
3774
David Rientjesd773ed62007-10-16 23:26:01 -07003775 ret = __zone_reclaim(zone, gfp_mask, order);
Johannes Weiner57054652014-10-09 15:28:17 -07003776 clear_bit(ZONE_RECLAIM_LOCKED, &zone->flags);
David Rientjesd773ed62007-10-16 23:26:01 -07003777
Mel Gorman24cf725182009-06-16 15:33:23 -07003778 if (!ret)
3779 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3780
David Rientjesd773ed62007-10-16 23:26:01 -07003781 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003782}
Christoph Lameter9eeff232006-01-18 17:42:31 -08003783#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003784
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003785/*
3786 * page_evictable - test whether a page is evictable
3787 * @page: the page to test
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003788 *
3789 * Test whether page is evictable--i.e., should be placed on active/inactive
Hugh Dickins39b5f292012-10-08 16:33:18 -07003790 * lists vs unevictable list.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003791 *
3792 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003793 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07003794 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003795 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003796 */
Hugh Dickins39b5f292012-10-08 16:33:18 -07003797int page_evictable(struct page *page)
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003798{
Hugh Dickins39b5f292012-10-08 16:33:18 -07003799 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003800}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003801
Hugh Dickins850465792012-01-20 14:34:19 -08003802#ifdef CONFIG_SHMEM
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003803/**
Hugh Dickins24513262012-01-20 14:34:21 -08003804 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3805 * @pages: array of pages to check
3806 * @nr_pages: number of pages to check
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003807 *
Hugh Dickins24513262012-01-20 14:34:21 -08003808 * Checks pages for evictability and moves them to the appropriate lru list.
Hugh Dickins850465792012-01-20 14:34:19 -08003809 *
3810 * This function is only used for SysV IPC SHM_UNLOCK.
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003811 */
Hugh Dickins24513262012-01-20 14:34:21 -08003812void check_move_unevictable_pages(struct page **pages, int nr_pages)
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003813{
Johannes Weiner925b7672012-01-12 17:18:15 -08003814 struct lruvec *lruvec;
Hugh Dickins24513262012-01-20 14:34:21 -08003815 struct zone *zone = NULL;
3816 int pgscanned = 0;
3817 int pgrescued = 0;
3818 int i;
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003819
Hugh Dickins24513262012-01-20 14:34:21 -08003820 for (i = 0; i < nr_pages; i++) {
3821 struct page *page = pages[i];
3822 struct zone *pagezone;
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003823
Hugh Dickins24513262012-01-20 14:34:21 -08003824 pgscanned++;
3825 pagezone = page_zone(page);
3826 if (pagezone != zone) {
3827 if (zone)
3828 spin_unlock_irq(&zone->lru_lock);
3829 zone = pagezone;
3830 spin_lock_irq(&zone->lru_lock);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003831 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003832 lruvec = mem_cgroup_page_lruvec(page, zone);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003833
Hugh Dickins24513262012-01-20 14:34:21 -08003834 if (!PageLRU(page) || !PageUnevictable(page))
3835 continue;
3836
Hugh Dickins39b5f292012-10-08 16:33:18 -07003837 if (page_evictable(page)) {
Hugh Dickins24513262012-01-20 14:34:21 -08003838 enum lru_list lru = page_lru_base_type(page);
3839
Sasha Levin309381fea2014-01-23 15:52:54 -08003840 VM_BUG_ON_PAGE(PageActive(page), page);
Hugh Dickins24513262012-01-20 14:34:21 -08003841 ClearPageUnevictable(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003842 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3843 add_page_to_lru_list(page, lruvec, lru);
Hugh Dickins24513262012-01-20 14:34:21 -08003844 pgrescued++;
3845 }
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003846 }
Hugh Dickins24513262012-01-20 14:34:21 -08003847
3848 if (zone) {
3849 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3850 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3851 spin_unlock_irq(&zone->lru_lock);
3852 }
Hugh Dickins850465792012-01-20 14:34:19 -08003853}
3854#endif /* CONFIG_SHMEM */